paper research job

Have a language expert improve your writing

Run a free plagiarism check in 10 minutes, generate accurate citations for free.

• Knowledge Base
• Research paper

How to Write a Research Paper | A Beginner's Guide

A research paper is a piece of academic writing that provides analysis, interpretation, and argument based on in-depth independent research.

Research papers are similar to academic essays , but they are usually longer and more detailed assignments, designed to assess not only your writing skills but also your skills in scholarly research. Writing a research paper requires you to demonstrate a strong knowledge of your topic, engage with a variety of sources, and make an original contribution to the debate.

This step-by-step guide takes you through the entire writing process, from understanding your assignment to proofreading your final draft.

Table of contents

Understand the assignment, choose a research paper topic, conduct preliminary research, develop a thesis statement, create a research paper outline, write a first draft of the research paper, write the introduction, write a compelling body of text, write the conclusion, the second draft, the revision process, research paper checklist, free lecture slides.

Completing a research paper successfully means accomplishing the specific tasks set out for you. Before you start, make sure you thoroughly understanding the assignment task sheet:

• Read it carefully, looking for anything confusing you might need to clarify with your professor.
• Identify the assignment goal, deadline, length specifications, formatting, and submission method.
• Make a bulleted list of the key points, then go back and cross completed items off as you’re writing.

Carefully consider your timeframe and word limit: be realistic, and plan enough time to research, write, and edit.

Scribbr Citation Checker New

The AI-powered Citation Checker helps you avoid common mistakes such as:

• Missing commas and periods
• Incorrect usage of “et al.”
• Ampersands (&) in narrative citations
• Missing reference entries

There are many ways to generate an idea for a research paper, from brainstorming with pen and paper to talking it through with a fellow student or professor.

You can try free writing, which involves taking a broad topic and writing continuously for two or three minutes to identify absolutely anything relevant that could be interesting.

You can also gain inspiration from other research. The discussion or recommendations sections of research papers often include ideas for other specific topics that require further examination.

Once you have a broad subject area, narrow it down to choose a topic that interests you, m eets the criteria of your assignment, and i s possible to research. Aim for ideas that are both original and specific:

• A paper following the chronology of World War II would not be original or specific enough.
• A paper on the experience of Danish citizens living close to the German border during World War II would be specific and could be original enough.

Note any discussions that seem important to the topic, and try to find an issue that you can focus your paper around. Use a variety of sources , including journals, books, and reliable websites, to ensure you do not miss anything glaring.

Do not only verify the ideas you have in mind, but look for sources that contradict your point of view.

• Is there anything people seem to overlook in the sources you research?
• Are there any heated debates you can address?
• Do you have a unique take on your topic?
• Have there been some recent developments that build on the extant research?

In this stage, you might find it helpful to formulate some research questions to help guide you. To write research questions, try to finish the following sentence: “I want to know how/what/why…”

A thesis statement is a statement of your central argument — it establishes the purpose and position of your paper. If you started with a research question, the thesis statement should answer it. It should also show what evidence and reasoning you’ll use to support that answer.

The thesis statement should be concise, contentious, and coherent. That means it should briefly summarize your argument in a sentence or two, make a claim that requires further evidence or analysis, and make a coherent point that relates to every part of the paper.

You will probably revise and refine the thesis statement as you do more research, but it can serve as a guide throughout the writing process. Every paragraph should aim to support and develop this central claim.

What can proofreading do for your paper?

Scribbr editors not only correct grammar and spelling mistakes, but also strengthen your writing by making sure your paper is free of vague language, redundant words, and awkward phrasing.

See editing example

A research paper outline is essentially a list of the key topics, arguments, and evidence you want to include, divided into sections with headings so that you know roughly what the paper will look like before you start writing.

A structure outline can help make the writing process much more efficient, so it’s worth dedicating some time to create one.

Your first draft won’t be perfect — you can polish later on. Your priorities at this stage are as follows:

• Maintaining forward momentum — write now, perfect later.
• Paying attention to clear organization and logical ordering of paragraphs and sentences, which will help when you come to the second draft.
• Expressing your ideas as clearly as possible, so you know what you were trying to say when you come back to the text.

You do not need to start by writing the introduction. Begin where it feels most natural for you — some prefer to finish the most difficult sections first, while others choose to start with the easiest part. If you created an outline, use it as a map while you work.

Do not delete large sections of text. If you begin to dislike something you have written or find it doesn’t quite fit, move it to a different document, but don’t lose it completely — you never know if it might come in useful later.

Paragraph structure

Paragraphs are the basic building blocks of research papers. Each one should focus on a single claim or idea that helps to establish the overall argument or purpose of the paper.

Example paragraph

George Orwell’s 1946 essay “Politics and the English Language” has had an enduring impact on thought about the relationship between politics and language. This impact is particularly obvious in light of the various critical review articles that have recently referenced the essay. For example, consider Mark Falcoff’s 2009 article in The National Review Online, “The Perversion of Language; or, Orwell Revisited,” in which he analyzes several common words (“activist,” “civil-rights leader,” “diversity,” and more). Falcoff’s close analysis of the ambiguity built into political language intentionally mirrors Orwell’s own point-by-point analysis of the political language of his day. Even 63 years after its publication, Orwell’s essay is emulated by contemporary thinkers.

Citing sources

It’s also important to keep track of citations at this stage to avoid accidental plagiarism . Each time you use a source, make sure to take note of where the information came from.

You can use our free citation generators to automatically create citations and save your reference list as you go.

APA Citation Generator MLA Citation Generator

The research paper introduction should address three questions: What, why, and how? After finishing the introduction, the reader should know what the paper is about, why it is worth reading, and how you’ll build your arguments.

What? Be specific about the topic of the paper, introduce the background, and define key terms or concepts.

Why? This is the most important, but also the most difficult, part of the introduction. Try to provide brief answers to the following questions: What new material or insight are you offering? What important issues does your essay help define or answer?

How? To let the reader know what to expect from the rest of the paper, the introduction should include a “map” of what will be discussed, briefly presenting the key elements of the paper in chronological order.

The major struggle faced by most writers is how to organize the information presented in the paper, which is one reason an outline is so useful. However, remember that the outline is only a guide and, when writing, you can be flexible with the order in which the information and arguments are presented.

One way to stay on track is to use your thesis statement and topic sentences . Check:

• topic sentences against the thesis statement;
• topic sentences against each other, for similarities and logical ordering;
• and each sentence against the topic sentence of that paragraph.

Be aware of paragraphs that seem to cover the same things. If two paragraphs discuss something similar, they must approach that topic in different ways. Aim to create smooth transitions between sentences, paragraphs, and sections.

The research paper conclusion is designed to help your reader out of the paper’s argument, giving them a sense of finality.

Trace the course of the paper, emphasizing how it all comes together to prove your thesis statement. Give the paper a sense of finality by making sure the reader understands how you’ve settled the issues raised in the introduction.

You might also discuss the more general consequences of the argument, outline what the paper offers to future students of the topic, and suggest any questions the paper’s argument raises but cannot or does not try to answer.

You should not :

• Offer new arguments or essential information
• Take up any more space than necessary
• Begin with stock phrases that signal you are ending the paper (e.g. “In conclusion”)

There are four main considerations when it comes to the second draft.

• Check how your vision of the paper lines up with the first draft and, more importantly, that your paper still answers the assignment.
• Identify any assumptions that might require (more substantial) justification, keeping your reader’s perspective foremost in mind. Remove these points if you cannot substantiate them further.
• Be open to rearranging your ideas. Check whether any sections feel out of place and whether your ideas could be better organized.
• If you find that old ideas do not fit as well as you anticipated, you should cut them out or condense them. You might also find that new and well-suited ideas occurred to you during the writing of the first draft — now is the time to make them part of the paper.

The goal during the revision and proofreading process is to ensure you have completed all the necessary tasks and that the paper is as well-articulated as possible.

Global concerns

• Confirm that your paper completes every task specified in your assignment sheet.
• Check for logical organization and flow of paragraphs.
• Check paragraphs against the introduction and thesis statement.

Fine-grained details

Check the content of each paragraph, making sure that:

• each sentence helps support the topic sentence.
• no unnecessary or irrelevant information is present.
• all technical terms your audience might not know are identified.

Next, think about sentence structure , grammatical errors, and formatting . Check that you have correctly used transition words and phrases to show the connections between your ideas. Look for typos, cut unnecessary words, and check for consistency in aspects such as heading formatting and spellings .

Finally, you need to make sure your paper is correctly formatted according to the rules of the citation style you are using. For example, you might need to include an MLA heading  or create an APA title page .

Scribbr’s professional editors can help with the revision process with our award-winning proofreading services.

Discover our paper editing service

Checklist: Research paper

I have followed all instructions in the assignment sheet.

My introduction presents my topic in an engaging way and provides necessary background information.

My introduction presents a clear, focused research problem and/or thesis statement .

My paper is logically organized using paragraphs and (if relevant) section headings .

Each paragraph is clearly focused on one central idea, expressed in a clear topic sentence .

Each paragraph is relevant to my research problem or thesis statement.

I have used appropriate transitions  to clarify the connections between sections, paragraphs, and sentences.

My conclusion provides a concise answer to the research question or emphasizes how the thesis has been supported.

My conclusion shows how my research has contributed to knowledge or understanding of my topic.

My conclusion does not present any new points or information essential to my argument.

I have provided an in-text citation every time I refer to ideas or information from a source.

I have included a reference list at the end of my paper, consistently formatted according to a specific citation style .

I have thoroughly revised my paper and addressed any feedback from my professor or supervisor.

I have followed all formatting guidelines (page numbers, headers, spacing, etc.).

You've written a great paper. Make sure it's perfect with the help of a Scribbr editor!

Open Google Slides Download PowerPoint

Is this article helpful?

Other students also liked.

• Writing a Research Paper Introduction | Step-by-Step Guide
• Writing a Research Paper Conclusion | Step-by-Step Guide
• Research Paper Format | APA, MLA, & Chicago Templates

More interesting articles

• Academic Paragraph Structure | Step-by-Step Guide & Examples
• Checklist: Writing a Great Research Paper
• How to Create a Structured Research Paper Outline | Example
• How to Write a Discussion Section | Tips & Examples
• How to Write Recommendations in Research | Examples & Tips
• How to Write Topic Sentences | 4 Steps, Examples & Purpose
• Research Paper Appendix | Example & Templates
• Research Paper Damage Control | Managing a Broken Argument
• What Is a Theoretical Framework? | Guide to Organizing

What is your plagiarism score?

We use cookies to enhance our website for you. Proceed if you agree to this policy or learn more about it.

• Essay Database >
• Essays Samples >
• Essay Types >
• Research Paper Example

Job Research Papers Samples For Students

1477 samples of this type

Regardless of how high you rate your writing abilities, it's always a worthy idea to check out an expertly written Research Paper example, especially when you're handling a sophisticated Job topic. This is exactly the case when WowEssays.com directory of sample Research Papers on Job will prove useful. Whether you need to think up a fresh and meaningful Job Research Paper topic or look into the paper's structure or formatting peculiarities, our samples will provide you with the required material.

Another activity area of our write my paper website is providing practical writing support to students working on Job Research Papers. Research help, editing, proofreading, formatting, plagiarism check, or even crafting fully original model Job papers upon your request – we can do that all! Place an order and buy a research paper now.

General Information On Dental Implant Treatment Essay

[Author Name] [Author Affiliation]

Good Example Of Research Paper On Police Discretion

Good research paper on the profession of psychiatrist.

Don't waste your time searching for a sample.

Get your research paper done by professional writers!

Just from $10/page

Good Research Paper About Myth About Immigrants Taking Away Jobs From Americans

The profession of psychiatrist research papers example, good example of research paper on occupational injuries and illnesses among emergency medical services (ems) workers., good example of research paper on generational diversity in the workforce.

[Institution Title]

Motivational Practice Research Paper Examples

Research paper on the strategic decision of integrating risk management in compensational level of the job, the occupation of antofagasta research paper examples, free unions in the us research paper sample.

Labor unions have been around in the United State of America since the Industrial Revolution but they did not become a force in this nation until the early 1900s. The major labor unions did not appear until the early 1930s but the earliest recorded labor dispute occurred in 1768 in New York.

Good Example Of Research Paper On Human Resource Management And The Healthcare Industry

Good example of human resource management research paper, research paper on haiti research paper, haiti: a history research paper examples, example of the effects of education on crime research paper, good example of research paper on free labor market and talent differences, employee turnover research papers examples, good example of research paper on a study of the effectiveness of safety measures for reducing accidents at the nursing home work place.

One major problem many organizations face is the occurrence of accidents involving trips, falls, and slips of employees. A rise in the injury rates is due to the lack of relevant and essential safety measures to ensure safety in the workplace. The study seeks to determine the presence and effectiveness of safety measures in reducing accidents at the nursing home workplace. It focuses on the occurrence of accidents among employees at one nursing home. The study examines the causes of slips, trips and falls (STF) accidents, the effectiveness of existing safety measures, and improvements that can occur.

Good Example Of Student Debt Crisis Research Paper

Hr white paper research paper samples, sample research paper on stockholders, job analysis research papers example, good the effects of unemployment on the economy research paper example, free nursing theory: caring theory research paper sample, good recycling, can it help the economy research paper example, should long term welfare recipients have a deadline date research paper samples, free research paper on role of women in honduras, free research paper about leadership and the officename:institutional affiliation, john mclane research paper, supervisory skills research paper example, sample research paper on more or less public investment, political communication during the stimulus bill debate research paper example, the fiscal and monetary policy and economic fluctuations research paper samples, labor force stats from the current population survey (2003 to 2013) research papers examples.

Unemployment in the United StatesUnemployment in the United States talks about the reasons and measures of U.S. unemployment and techniques for lessening it. Job creation and unemployment are influenced by elements, for example, investment conditions, global competition, education, mechanization, and demographics. These elements can influence the amount of specialists, the term of unemployment, and compensation levels. One real answer for anticipating unemployment in the United States is to keep the economy developing and creating through a dynamic business cycle. Using cash in a social order is one of the most ideal approaches to keep the economy consistent.

Research Paper On Project Management Steps For Seetting Up Plant: Switch System For 1100cc Bike

Sample research paper on economics aspect of united kingdom, example of research paper on reduce poverty and inequality in america, good example of research paper on research summary and ethical consideration.

06 febuary, 2014

Canada Research Paper Examples

The country of Canada just like the United States has come up with several laws and regulations protecting its public. These laws are supposed to protect the citizens from exploitation and to ensure equity and justice among the citizens. The United States government has throughout the years come up with laws and regulations that are meant to protect, preserve and guide it citizens in everything that they do. The labor sector has come up with laws that govern the eligibility of people to work within the country.

Climate Change In Political Discourse Research Paper Example

Free memorandum research paper sample, free undocumented workers research paper example, good unemployment in saudi arabia research paper example, example of reasons offered by market oriented and embeddedness in the u.s labor market research paper, free instructors name research paper example, tutorsubject research papers example, designing a compensation package for expatriates research paper samples, good research paper about homelessness, research paper on him, recycle and global economy research paper example, good work from house research paper example, example of research paper on labor supply and leisure choice, good example of course project - credentials research paper, free comp and coding management research paper sample, feminism in the united states research paper, other, public administration research papers example, education: an investment in your future research paper example, answers to questions research paper examples, good example of research paper on discrimination against women in the workplace, good example of research paper on elements of a quantitative research article, recent business conflict-past 36 months research paper example, organizational ethics: cirque du soleil research paper sample, good example of what is healthcare administrative partners research paper.

Research Paper

Good Research Paper On What We Can To Do To Make Orange County Prisons More Rehabilitative

The google bus conflict: functionalism research paper examples, developing an evaluation plan research paper examples, portrait of teresa research paper examples, good example of developing an implementation plan research paper, a reflection on education of deaf in residential schools research papers examples, racial stereotypes, cultural differences, & employment segregation research paper sample, sample research paper on slavery, immigration in the united states from mexico research papers example, example of the american economy in 2014 research paper.

Introduction 2 The Fiscal Policy 3 The Monetary Policy 4 Conclusion 4

Good Example Of Research Paper On How Haiti's Past Affected Haiti's Present

Gender diversity in the workplace research paper, free research paper on cancer-related illnesses and deaths in the state of wisconsin: cervical cancer in, good research paper about elizabeth angeli is now at department of english, purdue university.

For example Elizabeth L. Angeli, Department of Psychology, State University.

This research was supported in part by a grant from the Sample Grant Program. Angeli, Department of English, Purdue University, West Lafayette, IN 55555 The Open System Interconnect (OSI) model is a set of layers that provide the levels in which the information (or Network packets) takes different forms while travelling from one host to another host.

- Application Layer - Presentation Layer - Session Layer - Transport Layer - Network Layer - Data Link Layer - Physical Layer (networks?, 2009)

Research Paper On How Is The Computer Industry Gendered

Free operational human resource management research paper sample.

Performance evaluation is very important aspect of strategic human resource management in contemporary corporate management. Setting targets and assigning functions to staff in an organization without proper machinery in place to assess how these assigned tasks are executed will lead to a gap in effective management control and the unified effort at meeting the organization’s objectives and long term goals.

Good Research Paper About Breach Of Confidentiality

Hrm and business strategies alignment for marriott research paper, psychological considerations and sociological effects of 3-d printing technology research paper, good research paper on budgetary analysis, good research paper on agencys law and ethics of hiring a diverse workforce, example of musc project proposal research paper, free research paper on future of nursing, whistle blowing research paper sample, research paper on customer relationship management, free research paper on criminal law: final assessment, good research paper on busn311 - quantitative methods and analysis.

This paper presents the findings of three regression analysis performed on American InterContinental University’s survey data. They include regression analysis for intrinsic satisfaction versus benefits, extrinsic satisfaction versus benefits and overall satisfaction versus benefits. The coefficient of determination for each analysis was found to be 0.0042629, 0.000018508, and 0.0025798 respectively. In each case the coefficient of correlation was found to be less than 0.1. Meaning, there is very weak relationship between benefits and job satisfaction.

Planning And Leadership Research Paper Examples

Legalization of marijuana for profits in america research papers examples, good research paper on emotional intelligence: best practices, political issues relate with information system and technology research paper example, we've come a long way, baby: tips for hiring women in the 1940's research paper samples, good reflections on the necessity of college research paper example.

.001[Your Name]

Alignment Of HRM And Business Strategy Research Paper Sample

Academic research paper on facts.

Secretary of Labor v. Cities Service Oil Co.and theOccupational Safety & Health Review Commission

Government Assistance (Welfare) In The US Research Paper Examples

The natural rate of unemployment research paper examples, john lewis gaddis and the cold war research paper examples, air quality regulations in california research paper examples, research paper on the outlook for teachers: job analysis, three careers review research paper examples, research paper on physician assistant profession in 2014 and beyond, research paper on eeo law part ii, exploring two different theories of motivation research paper sample, research paper on determine the leadership styles and qualities used by both google management and, does an increase in minimum wage result in a higher unemployment rate research paper sample, the evolution of leadership strategies research paper examples, example of the agency and unions research paper.

The Administration for Children and Families (AFC) in the United States which falls under the Department of Health and Human Services has the responsibility of the common welfare as well as the financial well being of individuals, families, children and communities. Among its objectives, is to advance the health and wealth of families and children by providing resources, partnership and leadership for effective transfer of human services. It therefore aims at advancing hope and opportunities for persons in need of personal and social services.

Example Of International Human Resource Research Paper

Example of research paper on george orwell, research paper on probation in the state of georgia, research paper on nursing informatics competencies, example of education legislation research paper.

Technical education in California offers a sequence of programs which combine basic academic knowledge and professional knowledge to enable students get careers after secondary education. The department that is responsible for this program is the career technical Education department. The state education board uses the funds derived from the Perkins act to fund the courses. Perkins funds are available upon application and after satisfying the department of education that the programs are in line with the motives of the fund.

Effects Of Burnout Research Paper Examples

Research paper on typical hr structure of a school.

For the purpose of this discussion, the organizational set-up for a typical school shall be examined. The organizational chart for a school is shown below. Figure 1 Organizational Chart for a School

Four (4) critical positions from this organization are considered, the school principal, the assistant school principal, the classroom teacher, and the school nurse. Job descriptions and job guides are based on the US State Board of Education descriptions. Job descriptions for similar positions in a different educational institution may vary. Job guides for each position are listed below.

Example Of Governments Stopping Poverty Research Paper

Motivation case study research paper examples, link with motivation research paper.

DQ week 2 Three

Job attitudes have a direct link to the motivation levels in any organizations. Motivation influences employee behaviour. It is expected that highly motivated employees will be efficient and effective while low motivation will to dismal performance by employees. Employee motivation level also influences their interpersonal relations, as well as adherence to organizational policies.

Research Paper On Self-Assessment

DQ WK 2 Two

Likert-type scales are used in determining the level of an employee’s attitude. This method of determining job attitude is believed to be the best by some psychologists because of the wide range of information about an employee. Other methods of ranking employee attitude could be offering simple and clear ranking methods. However, the Likert scale stands out as one that gives the correct rank with accuracy. Likert scales do not measure job attitude because they use ordinal information.

Free Research Paper On Important Job Attitudes

DQ Week Two One Introduction There are some job attitudes that are outstanding and can be used to determine if an individual is a good employee or not. Therefore, some job attitudes are more important than others because of the information they help us understand. The other job attitudes are less important because they do not offer adequate information to enable to I-O psychologists assess employees.

Example Of Healthcare Policy Papers

Free research paper on racism is not good for people nor is it good for society, research paper on job attitudes matrix, research paper on should the trade embargo against cuba continue or not, measurement research paper sample.

DQ WK One 2 Introduction There are various kinds of job attitudes used in I/O psychology today. These attitudes are measurable for purposes of determining the positivity or negativity of employees towards the job. One such job attitude is the perceived organizational support (POS). This job attitude focuses on the level to which members of an organization feel that the organization cares and caters for their welfare.

Similarities Research Paper Examples

DQ WK 1 One

Job satisfaction and organizational commitment are two concepts that have similarities but remain different. Job satisfaction can be described as the level of contentedness of an employee about their job. On the other hand, organizational commitment is the level of psychological linkage of an individual to an organization they work. Therefore, job satisfaction is at an individual level where one enjoys the duties he or she performs while organizational commitment is the loyalty of an individual to an organization and the work they do.

Free Research Paper On Is China Destined, In The 21st Century, To Replace The USA As The World's Dominant

Research paper on evaluation of agencys public personnel administration.

The Administration For Children And Families (AFC) endeavors to deliver quality services to the targeted youths, children and families. These services range from Child support, welfare, foster care, adoption assistance to overall child care programs. In order to achieve this crucial objective, the AFC is composed of an organized human resource management structure which makes it possible for this organization to deliver their services to the concerned parties. In this section of my assignment, I will major my focus on the organization design of AFC in regard to its human resource.

Example Of SEC Fillings Since 2010 Research Paper

The company provided all documents as per SEC requirements. In its submissions it stated the changes share ownership, the trading price of shares and all other information which is deemed to affect the price of company shares. The company’s market share prices have been rising steadily over the last three years (Starbucks Coffee Company, n.d).

Research Paper About Untouchability: Kabir’s and B. R. Ambedkar’s Takes

The value of motherhood research paper, the effect of internet connectivity in job search research paper example, maslow hierarchy of needs research paper sample, generation jobless unemployment research paper example, employment-at-will doctrine research paper examples, example of effectiveness and ethics in leadership research paper, professional development planning research paper, example of analyzing the different measurements used in the selection process research paper, japan's history and present government research paper examples, research paper on identify and briefly describe the constraints involved in developing a selection, osha emergency action plans and fire prevention plans research paper sample, mw brand research paper example, captain strong research paper example.

There are various motivational strategies which Captain Strong could consider adopting. The first strategy she could consider adopting would be empowerment. Empowering the patrol division officers would entail giving them the trust, authority, autonomy and encouragement to complete a task. This will be an important strategy given the fact that among the complaints of the officers is the fact that they hardly ever get to complete their tasks due to the many orders they receive. In short, they lack the necessary autonomy and encouragement to complete their tasks.

Quality Improvement Assignment Research Paper Examples

Research paper on political support, research paper on automatic data processing, labour force participation rate research paper, research paper on job criteria.

Apparently, many people see money as the primary determinant of whether or not a person will take up a job. This is a serious misconception as people are different and hold different views on the same. While others prioritize money, others will look at such things as the working hours, the chances to travel and so on. Personally, I would prioritize job security, work-life balance and room for growth as my first three determinants. Faced with many offers with equal pay, I would consider these three criteria for various reasons discussed below.

The Company Nike Research Paper Examples

Colin powell leadership traits research paper examples, what factors affect the labor market research paper sample, research paper on fred hutchinson, winners and losers research paper examples, scientific management: is taylorism still practiced in workplaces research paper sample, free research paper on job satisfaction, research paper on international business and globalization, free research paper on should public servants have the right to strike, low wage living research paper sample, example of change and culture case 11 research paper, free research paper on social inequality, selection techniques research paper example.

The selection criterion for different workers always varies depending on several factors. The criterion for selection of top executive employees is totally different from the criterion used to select middle level managers and that of junior employees. These criteria of selection are selected based on the rank of employees, responsibilities of the job and the selection method that would help identify the necessary skills and personality in potential employees for each post.

How Can We Make School Better For Today Research Paper Examples

Example of research paper on oppression of the mentally ill, importance of interviews research paper examples.

The interview process is important in any hiring organization because it helps the organization identify the most suitable candidates for the job. The organization has to consider several factors before settling on either a structured or unstructured interview process. The questions asked in both kinds of interviews differ in several ways.

Employer Information Research Paper Examples

Research paper on the nayars of india and their subsistence, employee selection and development proposal research paper sample.

Gilliland-Moore Wines, Annetta Cherne, Introduction

Gilliland-Moore Wines has undergone rapid growth recently. The company has expanded its supply base to the national level. Therefore, it has to select and develop new employees to help ease the job burden created by the company’s expansion. The company must conduct a research to determine a suitable method of selecting and developing competent employees.

Importance Of Choosing Appropriate Methods Research Paper Example

Success of any organization is based on its ability to choose suitable methods to select, hire and develop potential employees. This is because the kind of employees a company develops or hires are the ones that determine the organization’s productivity. Therefore, organization must emphasize on comprehensive methods of selecting, hiring and developing potential employees.

Research Paper On Job Position Office Manager Pinnacle Mountain State Park

Based on the duties and qualifications, this job position is equivalent to Assistant Retail Store Manager – Museum (Salary.com, 2012). The median salary for such a position in the U.S. is $36,491. Described hereunder is the compensation and benefits plan for the employee in the above position according to Salary.com, 2012.

Amount ($) Percentage Base Salary

Social Security

Example Of Formal Report Outsourcing And Its Alternatives Research Paper

This paper will examine the current trends in outsourcing and its alternatives an organization. The focus of the paper will be on the shift in outsourcing destinations and the motivation behind the changes. Onshore and offshore respondents will be interviewed to establish a trend on both outsourcing destinations. The conclusions will be based on these results and a projection of future trends.

Current Rehabilitation Issues Research Paper Examples

Social rehabilitation in criminal law is the reforming of a criminal into a law abiding individual, fit to live in harmony with other citizens. Total social rehabilitation can be achieved, even though it is a difficult for most offenders of the law. Total social rehabilitation requires the services of psychologists and social counselors. The offender would be guided on how to be a law abiding citizen and how to live harmoniously with other people in society. If offenders go through a thorough rehabilitation process, they can renew themselves completely to fit into society without any problem.

Example Of Some Of The Standards Used By Consultants In Criterion Development Include A Broad Research Paper

The development of a selection standard is important because it is a critical tool for organizational consultants in the selection process. The criteria’s primary objective is to guide the consultants determine the candidates that qualify for the position available. The selection standards must be determined before the advertisement of a vacancy. The criteria should be clearly defined and have relevancy to the position. This will make it easy to evaluate candidates because it is defensible, understandable and in alignment with the needs of the department or organization (Schultz & Schultz, 2010).

Comprehensiveness Research Paper Examples

A competence model is a standard of describing and analyzing the range, and types of skills, knowledge and abilities an organization seeks from its existing and potential employees in order to have a competitive advantage. On the other hand, traditional job analysis techniques are the standards or criterions developed to determine the job descriptions hence help in determining the people to be hired for the job and how the job is done. Traditional techniques of job analysis are more important than competency model because of several reasons (Schultz & Schultz, 2010).

Human Resource Planning Research Paper Example

An elaborated job analysis can be a basis of good personnel management. The employer and the employee need to understand the responsibilities and duties employees should do. Job analysis can be used to enhance understanding since it links each unit of work to other units. There are five common uses of job analysis.

Nonparametric Tests Research Paper Sample

Nonparametric test are used by researchers when the collected data is categorical or ranked rather than quantified numerically. Unlike parametric tests, nonparametric test does not require certain assumptions, which make it more flexible. For, example the assumption that the data is normally distributed is relaxed when conducting nonparametric tests. Nonparametric tests are commonly used in psychological research because, the data collected often entails ranked responses; the data collected is non-numerical and is not normally distributed.

Research Paper On Selecting A Statistical Test

Research paper on strong and weak governors, life stages research paper examples, research paper on bc aboriginal title, research paper on arab-americans in the united states.

Since the September 11th terrorist attacks on the United States, civil liberties have significantly deteriorated for all the American citizens. Precisely, legislations passed immediately after the attacks have undermined Arab-American perceptions of safety in the nation. This paper seeks to focus on discrimination in employment of Arabs-American, since the attacks evident by media reports regarding the increase in the issue. Henceforth, the paper will come up with recommendations for US employers to curb incidences of discrimination.

Free Research Paper On Analyzing Impact Of Number Of Years Of Employment On Annual Wages

Knowledge transfer research paper, skin cancer in men and women research paper examples.

Contrary to what most people believe, American men are more prone to skin cancer than women (Voiland, 2008). In fact, according to statistics, more men die due to skin cancer than women every year and there are reasons behind this fact. This paper will identify some of the most common reasons why men are more prone to skin cancer than women using strong evidences from journals and publications synthesized over the internet and from books.

Example Of Research Paper On Leadership And Change For Nintendo

Stereotyping in america research paper.

Password recovery email has been sent to [email protected]

Use your new password to log in

You are not register!

By clicking Register, you agree to our Terms of Service and that you have read our Privacy Policy .

Now you can download documents directly to your device!

Check your email! An email with your password has already been sent to you! Now you can download documents directly to your device.

or Use the QR code to Save this Paper to Your Phone

The sample is NOT original!

Short on a deadline?

Don't waste time. Get help with 11% off using code - GETWOWED

No, thanks! I'm fine with missing my deadline

Help | Advanced Search

Computer Science > Computation and Language

Title: code llama: open foundation models for code.

Abstract: We release Code Llama, a family of large language models for code based on Llama 2 providing state-of-the-art performance among open models, infilling capabilities, support for large input contexts, and zero-shot instruction following ability for programming tasks. We provide multiple flavors to cover a wide range of applications: foundation models (Code Llama), Python specializations (Code Llama - Python), and instruction-following models (Code Llama - Instruct) with 7B, 13B and 34B parameters each. All models are trained on sequences of 16k tokens and show improvements on inputs with up to 100k tokens. 7B and 13B Code Llama and Code Llama - Instruct variants support infilling based on surrounding content. Code Llama reaches state-of-the-art performance among open models on several code benchmarks, with scores of up to 53% and 55% on HumanEval and MBPP, respectively. Notably, Code Llama - Python 7B outperforms Llama 2 70B on HumanEval and MBPP, and all our models outperform every other publicly available model on MultiPL-E. We release Code Llama under a permissive license that allows for both research and commercial use.

Submission history

• Download a PDF of the paper titled Code Llama: Open Foundation Models for Code, by Baptiste Rozi\`ere and 24 other authors PDF
• Other formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open Access
• Published: 30 August 2023

Champion-level drone racing using deep reinforcement learning

• Elia Kaufmann   ORCID: orcid.org/0000-0001-6094-5901 1 ,
• Leonard Bauersfeld 1 ,
• Antonio Loquercio   ORCID: orcid.org/0000-0002-8410-3933 1 ,
• Matthias Müller 2 ,
• Vladlen Koltun 3 &
• Davide Scaramuzza   ORCID: orcid.org/0000-0002-3831-6778 1  

Nature volume  620 ,  pages 982–987 ( 2023 ) Cite this article

28k Accesses

1015 Altmetric

Metrics details

• Aerospace engineering
• Computer science
• Electrical and electronic engineering
• Mechanical engineering

First-person view (FPV) drone racing is a televised sport in which professional competitors pilot high-speed aircraft through a 3D circuit. Each pilot sees the environment from the perspective of their drone by means of video streamed from an onboard camera. Reaching the level of professional pilots with an autonomous drone is challenging because the robot needs to fly at its physical limits while estimating its speed and location in the circuit exclusively from onboard sensors 1 . Here we introduce Swift, an autonomous system that can race physical vehicles at the level of the human world champions. The system combines deep reinforcement learning (RL) in simulation with data collected in the physical world. Swift competed against three human champions, including the world champions of two international leagues, in real-world head-to-head races. Swift won several races against each of the human champions and demonstrated the fastest recorded race time. This work represents a milestone for mobile robotics and machine intelligence 2 , which may inspire the deployment of hybrid learning-based solutions in other physical systems.

Deep RL 3 has enabled some recent advances in artificial intelligence. Policies trained with deep RL have outperformed humans in complex competitive games, including Atari 4 , 5 , 6 , Go 5 , 7 , 8 , 9 , chess 5 , 9 , StarCraft 10 , Dota 2 (ref.  11 ) and Gran Turismo 12 , 13 . These impressive demonstrations of the capabilities of machine intelligence have primarily been limited to simulation and board-game environments, which support policy search in an exact replica of the testing conditions. Overcoming this limitation and demonstrating champion-level performance in physical competitions is a long-standing problem in autonomous mobile robotics and artificial intelligence 14 , 15 , 16 .

FPV drone racing is a televised sport in which highly trained human pilots push aerial vehicles to their physical limits in high-speed agile manoeuvres (Fig. 1a ). The vehicles used in FPV racing are quadcopters, which are among the most agile machines ever built (Fig. 1b ). During a race, the vehicles exert forces that surpass their own weight by a factor of five or more, reaching speeds of more than 100 km h −1 and accelerations several times that of gravity, even in confined spaces. Each vehicle is remotely controlled by a human pilot who wears a headset showing a video stream from an onboard camera, creating an immersive ‘first-person-view’ experience (Fig. 1c ).

a , Swift (blue) races head-to-head against Alex Vanover, the 2019 Drone Racing League world champion (red). The track comprises seven square gates that must be passed in order in each lap. To win a race, a competitor has to complete three consecutive laps before its opponent. b , A close-up view of Swift, illuminated with blue LEDs, and a human-piloted drone, illuminated with red LEDs. The autonomous drones used in this work rely only on onboard sensory measurements, with no support from external infrastructure, such as motion-capture systems. c , From left to right: Thomas Bitmatta, Marvin Schaepper and Alex Vanover racing their drones through the track. Each pilot wears a headset that shows a video stream transmitted in real time from a camera aboard their aircraft. The headsets provide an immersive ‘first-person-view’ experience. c , Photo by Regina Sablotny.

Attempts to create autonomous systems that reach the performance of human pilots date back to the first autonomous drone racing competition in 2016 (ref.  17 ). A series of innovations followed, including the use of deep networks to identify the next gate location 18 , 19 , 20 , transfer of racing policies from simulation to reality 21 , 22 and accounting for uncertainty in perception 23 , 24 . The 2019 AlphaPilot autonomous drone racing competition showcased some of the best research in the field 25 . However, the first two teams still took almost twice as long as a professional human pilot to complete the track 26 , 27 . More recently, autonomous systems have begun to reach expert human performance 28 , 29 , 30 . However, these works rely on near-perfect state estimation provided by an external motion-capture system. This makes the comparison with human pilots unfair, as humans only have access to onboard observations from the drone.

In this article, we describe Swift, an autonomous system that can race a quadrotor at the level of human world champions using only onboard sensors and computation. Swift consists of two key modules: (1) a perception system that translates high-dimensional visual and inertial information into a low-dimensional representation and (2) a control policy that ingests the low-dimensional representation produced by the perception system and produces control commands.

The control policy is represented by a feedforward neural network and is trained in simulation using model-free on-policy deep RL 31 . To bridge discrepancies in sensing and dynamics between simulation and the physical world, we make use of non-parametric empirical noise models estimated from data collected on the physical system. These empirical noise models have proved to be instrumental for successful transfer of the control policy from simulation to reality.

We evaluate Swift on a physical track designed by a professional drone-racing pilot (Fig. 1a ). The track comprises seven square gates arranged in a volume of 30 × 30 × 8 m, forming a lap of 75 m in length. Swift raced this track against three human champions: Alex Vanover, the 2019 Drone Racing League world champion, Thomas Bitmatta, two-time MultiGP International Open World Cup champion, and Marvin Schaepper, three-time Swiss national champion. The quadrotors used by Swift and by the human pilots have the same weight, shape and propulsion. They are similar to drones used in international competitions.

The human pilots were given one week of practice on the race track. After this week of practice, each pilot competed against Swift in several head-to-head races (Fig. 1a,b ). In each head-to-head race, two drones (one controlled by a human pilot and one controlled by Swift) start from a podium. The race is set off by an acoustic signal. The first vehicle that completes three full laps through the track, passing all gates in the correct order in each lap, wins the race.

Swift won several races against each of the human pilots and achieved the fastest race time recorded during the events. Our work marks the first time, to our knowledge, that an autonomous mobile robot achieved world-champion-level performance in a real-world competitive sport.

The Swift system

Swift uses a combination of learning-based and traditional algorithms to map onboard sensory readings to control commands. This mapping comprises two parts: (1) an observation policy, which distils high-dimensional visual and inertial information into a task-specific low-dimensional encoding, and (2) a control policy that transforms the encoding into commands for the drone. A schematic overview of the system is shown in Fig. 2 .

Swift consists of two key modules: a perception system that translates visual and inertial information into a low-dimensional state observation and a control policy that maps this state observation to control commands. Control commands specify desired collective thrust and body rates, the same control modality that the human pilots use. a , The perception system consists of a VIO module that computes a metric estimate of the drone state from camera images and high-frequency measurements obtained by an inertial measurement unit (IMU). The VIO estimate is coupled with a neural network that detects the corners of racing gates in the image stream. The corner detections are mapped to a 3D pose and fused with the VIO estimate using a Kalman filter. b , We use model-free on-policy deep RL to train the control policy in simulation. During training, the policy maximizes a reward that combines progress towards the centre of the next racing gate with a perception objective to keep the next gate in the field of view of the camera. To transfer the racing policy from simulation to the physical world, we augment the simulation with data-driven residual models of the vehicle’s perception and dynamics. These residual models are identified from real-world experience collected on the race track. MLP, multilayer perceptron.

The observation policy consists of a visual–inertial estimator 32 , 33 that operates together with a gate detector 26 , which is a convolutional neural network that detects the racing gates in the onboard images. Detected gates are then used to estimate the global position and orientation of the drone along the race track. This is done by a camera-resectioning algorithm 34 in combination with a map of the track. The estimate of the global pose obtained from the gate detector is then combined with the estimate from the visual–inertial estimator by means of a Kalman filter, resulting in a more accurate representation of the robot’s state. The control policy, represented by a two-layer perceptron, maps the output of the Kalman filter to control commands for the aircraft. The policy is trained using on-policy model-free deep RL 31 in simulation. During training, the policy maximizes a reward that combines progress towards the next racing gate 35 with a perception objective that rewards keeping the next gate in the field of view of the camera. Seeing the next gate is rewarded because it increases the accuracy of pose estimation.

Optimizing a policy purely in simulation yields poor performance on physical hardware if the discrepancies between simulation and reality are not mitigated. The discrepancies are caused primarily by two factors: (1) the difference between simulated and real dynamics and (2) the noisy estimation of the robot’s state by the observation policy when provided with real sensory data. We mitigate these discrepancies by collecting a small amount of data in the real world and using this data to increase the realism of the simulator.

Specifically, we record onboard sensory observations from the robot together with highly accurate pose estimates from a motion-capture system while the drone is racing through the track. During this data-collection phase, the robot is controlled by a policy trained in simulation that operates on the pose estimates provided by the motion-capture system. The recorded data allow to identify the characteristic failure modes of perception and dynamics observed through the race track. These intricacies of failing perception and unmodelled dynamics are dependent on the environment, platform, track and sensors. The perception and dynamics residuals are modelled using Gaussian processes 36 and k -nearest-neighbour regression, respectively. The motivation behind this choice is that we empirically found perception residuals to be stochastic and dynamics residuals to be largely deterministic (Extended Data Fig. 1 ). These residual models are integrated into the simulation and the racing policy is fine-tuned in this augmented simulation. This approach is related to the empirical actuator models used for simulation-to-reality transfer in ref.  37 but further incorporates empirical modelling of the perception system and also accounts for the stochasticity in the estimate of the platform state.

We ablate each component of Swift in controlled experiments reported in the extended data. Also, we compare against recent work that tackles the task of autonomous drone racing with traditional methods, including trajectory planning and model predictive control (MPC). Although such approaches achieve comparable or even superior performance to our approach in idealized conditions, such as simplified dynamics and perfect knowledge of the robot’s state, their performance collapses when their assumptions are violated. We find that approaches that rely on precomputed paths 28 , 29 are particularly sensitive to noisy perception and dynamics. No traditional method has achieved competitive lap times compared with Swift or human world champions, even when provided with highly accurate state estimation from a motion-capture system. Detailed analysis is provided in the extended data.

The drone races take place on a track designed by an external world-class FPV pilot. The track features characteristic and challenging manoeuvres, such as a Split-S (Figs. 1a (top-right corner) and 4d ). Pilots are allowed to continue racing even after a crash, provided their vehicle is still able to fly. If both drones crash and cannot complete the track, the drone that proceeded farther along the track wins.

As shown in Fig. 3b , Swift wins 5 out of 9 races against A. Vanover, 4 out of 7 races against T. Bitmatta and 6 out of 9 races against M. Schaepper. Out of the 10 losses recorded for Swift, 40% were because of a collision with the opponent, 40% because of collision with a gate and 20% because of the drone being slower than the human pilot. Overall, Swift wins most races against each human pilot. Swift also achieves the fastest race time recorded, with a lead of half a second over the best time clocked by a human pilot (A. Vanover).

a , Lap-time results. We compare Swift against the human pilots in time-trial races. Lap times indicate best single lap times and best average times achieved in a heat of three consecutive laps. The reported statistics are computed over a dataset recorded during one week on the race track, which corresponds to 483 (115) data points for Swift, 331 (221) for A. Vanover, 469 (338) for T. Bitmatta and 345 (202) for M. Schaepper. The first number is the number of single laps and the second is the number of three consecutive laps. The dark points in each distribution correspond to laps flown in race conditions. b , Head-to-head results. We report the number of head-to-head races flown by each pilot, the number of wins and losses, as well as the win ratio.

Figure 4 and Extended Data Table 1d provide an analysis of the fastest lap flown by Swift and each human pilot. Although Swift is globally faster than all human pilots, it is not faster on all individual segments of the track (Extended Data Table 1 ). Swift is consistently faster at the start and in tight turns such as the split S. At the start, Swift has a lower reaction time, taking off from the podium, on average, 120 ms before human pilots. Also, it accelerates faster and reaches higher speeds going into the first gate (Extended Data Table 1d , segment 1). In sharp turns, as shown in Fig. 4c,d , Swift finds tighter manoeuvres. One hypothesis is that Swift optimizes trajectories on a longer timescale than human pilots. It is known that model-free RL can optimize long-term rewards through a value function 38 . Conversely, human pilots plan their motion on a shorter timescale, up to one gate into the future 39 . This is apparent, for example in the split S (Fig. 4b,d ), for which human pilots are faster in the beginning and at the end of the manoeuvre, but slower overall (Extended Data Table 1d , segment 3). Also, human pilots orient the aircraft to face the next gate earlier than Swift does (Fig. 4c,d ). We propose that human pilots are accustomed to keeping the upcoming gate in view, whereas Swift has learned to execute some manoeuvres while relying on other cues, such as inertial data and visual odometry against features in the surrounding environments. Overall, averaged over the entire track, the autonomous drone achieves the highest average speed, finds the shortest racing line and manages to maintain the aircraft closer to its actuation limits throughout the race, as indicated by the average thrust and power drawn (Extended Data Table 1d ).

a , Comparison of the fastest race of each pilot, illustrated by the time behind Swift. The time difference from the autonomous drone is computed as the time since it passed the same position on the track. Although Swift is globally faster than all human pilots, it is not necessarily faster on all individual segments of the track. b , Visualization of where the human pilots are faster (red) and slower (blue) compared with the autonomous drone. Swift is consistently faster at the start and in tight turns, such as the split S. c , Analysis of the manoeuvre after gate 2. Swift in blue, Vanover in red. Swift gains time against human pilots in this segment as it executes a tighter turn while maintaining comparable speed. d , Analysis of the split S manoeuvre. Swift in blue, Vanover in red. The split S is the most challenging segment in the race track, requiring a carefully coordinated roll and pitch motion that yields a descending half-loop through the two gates. Swift gains time against human pilots on this segment as it executes a tighter turn with less overshoot. e , Illustration of track segments used for analysis. Segment 1 is traversed once at the start, whereas segments 2–4 are traversed in each lap (three times over the course of a race).

We also compare the performance of Swift and the human champions in time trials (Fig. 3a ). In a time trial, a single pilot races the track, with the number of laps left to the discretion of the pilot. We accumulate time-trial data from the practice week and the races, including training runs (Fig. 3a , coloured) and laps flown in race conditions (Fig. 3a , black). For each contestant, we use more than 300 laps for computing statistics. The autonomous drone more consistently pushes for fast lap times, exhibiting lower mean and variance. Conversely, human pilots decide whether to push for speed on a lap-by-lap basis, yielding higher mean and variance in lap times, both during training and in the races. The ability to adapt the flight strategy allows human pilots to maintain a slower pace if they identify that they have a clear lead, so as to reduce the risk of a crash. The autonomous drone is unaware of its opponent and pushes for fastest expected completion time no matter what, potentially risking too much when in the lead and too little when trailing behind 40 .

FPV drone racing requires real-time decision-making based on noisy and incomplete sensory input from the physical environment. We have presented an autonomous physical system that achieves champion-level performance in this sport, reaching—and at times exceeding—the performance of human world champions. Our system has certain structural advantages over the human pilots. First, it makes use of inertial data from an onboard inertial measurement unit 32 . This is akin to the human vestibular system 41 , which is not used by the human pilots because they are not physically in the aircraft and do not feel the accelerations acting on it. Second, our system benefits from lower sensorimotor latency (40 ms for Swift versus an average of 220 ms for expert human pilots 39 ). On the other hand, the limited refresh rate of the camera used by Swift (30 Hz) can be considered a structural advantage for human pilots, whose cameras’ refresh rate is four times as fast (120 Hz), improving their reaction time 42 .

Human pilots are impressively robust: they can crash at full speed, and—if the hardware still functions—carry on flying and complete the track. Swift was not trained to recover after a crash. Human pilots are also robust to changes in environmental conditions, such as illumination, which can markedly alter the appearance of the track. By contrast, Swift’s perception system assumes that the appearance of the environment is consistent with what was observed during training. If this assumption fails, the system can fail. Robustness to appearance changes can be provided by training the gate detector and the residual observation model in a diverse set of conditions. Addressing these limitations could enable applying the presented approach in autonomous drone racing competitions in which access to the environment and the drone is limited 25 .

Notwithstanding the remaining limitations and the work ahead, the attainment by an autonomous mobile robot of world-champion-level performance in a popular physical sport is a milestone for robotics and machine intelligence. This work may inspire the deployment of hybrid learning-based solutions in other physical systems, such as autonomous ground vehicles, aircraft and personal robots, across a broad range of applications.

Quadrotor simulation

Quadrotor dynamics.

To enable large-scale training, we use a high-fidelity simulation of the quadrotor dynamics. This section briefly explains the simulation. The dynamics of the vehicle can be written as

in which ⊙ represents quaternion rotation, \({{\bf{p}}}_{{\mathcal{W}}{\mathcal{B}}},{{\bf{q}}}_{{\mathcal{W}}{\mathcal{B}}},{{\bf{v}}}_{{\mathcal{W}}}\) and \({{\boldsymbol{\omega }}}_{{\mathcal{B}}}\) denote the position, attitude quaternion, inertial velocity and body rates of the quadcopter, respectively. The motor time constant is k mot and the motor speeds Ω and Ω ss are the actual and steady-state motor speeds, respectively. The matrix J  is the inertia of the quadcopter and \({{\bf{g}}}_{{\mathcal{W}}}\) denotes the gravity vector. Two forces act on the quadrotor: the lift force f prop generated by the propellers and an aerodynamic force f aero that aggregates all other forces, such as aerodynamic drag, dynamic lift and induced drag. The torque is modelled as a sum of four components: the torque generated by the individual propeller thrusts τ prop , the yaw torque τ mot generated by a change in motor speed, an aerodynamic torque τ aero that accounts for various aerodynamic effects such as blade flapping and an inertial term τ iner . The individual components are given as

in which r P, i is the location of propeller i , expressed in the body frame, and f i and τ i are the forces and torques, respectively, generated by the i th propeller. The axis of rotation of the i th motor is denoted by ζ i , the combined inertia of the motor and propeller is J m+p and the derivative of the i th motor speed is \({\dot{\Omega }}_{i}\) . The individual propellers are modelled using a commonly used quadratic model, which assumes that the lift force and drag torque are proportional to the square of the propeller speed Ω i :

in which c l and c d denote the propeller lift and drag coefficients, respectively.

Aerodynamic forces and torques

The aerodynamic forces and torques are difficult to model with a first-principles approach. We thus use a data-driven model 43 . To maintain the low computational complexity required for large-scale RL training, a grey-box polynomial model is used rather than a neural network. The aerodynamic effects are assumed to primarily depend on the velocity \({{\bf{v}}}_{{\mathcal{B}}}\) (in the body frame) and the average squared motor speed \(\overline{{\Omega }^{2}}\) . The aerodynamic forces f x , f y and f z and torques τ x , τ y and τ z are estimated in the body frame. The quantities v x , v y and v z denote the three axial velocity components (in the body frame) and v x y denotes the speed in the ( x ,  y ) plane of the quadrotor. On the basis of insights from the underlying physical processes, linear and quadratic combinations of the individual terms are selected. For readability, the coefficients multiplying each summand have been omitted:

The respective coefficients are then identified from real-world flight data, in which motion capture is used to provide ground-truth forces and torque measurements. We use data from the race track, allowing the dynamics model to fit the track. This is akin to the human pilots’ training for days or weeks before the race on the specific track that they will be racing. In our case, the human pilots are given a week of practice on the same track ahead of the competition.

Betaflight low-level controller

To control the quadrotor, the neural network outputs collective thrust and body rates. This control signal is known to combine high agility with good robustness to simulation-to-reality transfer 44 . The predicted collective thrust and body rates are then processed by an onboard low-level controller that computes individual motor commands, which are subsequently translated into analogue voltage signals through an electronic speed controller (ESC) that controls the motors. On the physical vehicle, this low-level proportional–integral–derivative (PID) controller and ESC are implemented using the open-source Betaflight and BLHeli32 firmware 45 . In simulation, we use an accurate model of both the low-level controller and the motor speed controller.

Because the Betaflight PID controller has been optimized for human-piloted flight, it exhibits some peculiarities, which the simulation correctly captures: the reference for the D-term is constantly zero (pure damping), the I-term gets reset when the throttle is cut and, under motor thrust saturation, the body rate control is assigned priority (proportional downscaling of all motor signals to avoid saturation). The gains of the controller used for simulation have been identified from the detailed logs of the Betaflight controller’s internal states. The simulation can predict the individual motor commands with less than 1% error.

Battery model and ESC

The low-level controller converts the individual motor commands into a pulse-width modulation (PWM) signal and sends it to the ESC, which controls the motors. Because the ESC does not perform closed-loop control of the motor speeds, the steady-state motor speed Ω i ,ss for a given PWM motor command cmd i is a function of the battery voltage. Our simulation thus models the battery voltage using a grey-box battery model 46 that simulates the voltage based on instantaneous power consumption P mot :

The battery model 46 then simulates the battery voltage based on this power demand. Given the battery voltage U bat and the individual motor command u cmd, i , we use the mapping (again omitting the coefficients multiplying each summand)

to calculate the corresponding steady-state motor speed Ω i ,ss required for the dynamics simulation in equation ( 1 ). The coefficients have been identified from Betaflight logs containing measurements of all involved quantities. Together with the model of the low-level controller, this enables the simulator to correctly translate an action in the form of collective thrust and body rates to desired motor speeds Ω ss in equation ( 1 ).

Policy training

We train deep neural control policies that directly map observations o t in the form of platform state and next gate observation to control actions u t in the form of mass-normalized collective thrust and body rates 44 . The control policies are trained using model-free RL in simulation.

Training algorithm

Training is performed using proximal policy optimization 31 . This actor-critic approach requires jointly optimizing two neural networks during training: the policy network, which maps observations to actions, and the value network, which serves as the ‘critic’ and evaluates actions taken by the policy. After training, only the policy network is deployed on the robot.

Observations, actions and rewards

An observation \({{\bf{o}}}_{t}\in {{\mathbb{R}}}^{31}\) obtained from the environment at time t consists of: (1) an estimate of the current robot state; (2) the relative pose of the next gate to be passed on the track layout; and (3) the action applied in the previous step. Specifically, the estimate of the robot state contains the position of the platform, its velocity and attitude represented by a rotation matrix, resulting in a vector in \({{\mathbb{R}}}^{15}\) . Although the simulation uses quaternions internally, we use a rotation matrix to represent attitude to avoid ambiguities 47 . The relative pose of the next gate is encoded by providing the relative position of the four gate corners with respect to the vehicle, resulting in a vector in \({{\mathbb{R}}}^{12}\) . All observations are normalized before being passed to the network. Because the value network is only used during training time, it can access privileged information about the environment that is not accessible to the policy 48 . This privileged information is concatenated with other inputs to the policy network and contains the exact position, orientation and velocity of the robot.

For each observation o t , the policy network produces an action \({{\bf{a}}}_{t}\in {{\mathbb{R}}}^{4}\) in the form of desired mass-normalized collective thrust and body rates.

We use a dense shaped reward formulation to learn the task of perception-aware autonomous drone racing. The reward r t at time step t is given by

in which r prog rewards progress towards the next gate 35 , r perc encodes perception awareness by adjusting the attitude of the vehicle such that the optical axis of the camera points towards the centre of the next gate, r cmd rewards smooth actions and r crash is a binary penalty that is only active when colliding with a gate or when the platform leaves a predefined bounding box. If r crash is triggered, the training episode ends.

Specifically, the reward terms are

in which \({d}_{t}^{{\rm{Gate}}}\) denotes the distance from the centre of mass of the vehicle to the centre of the next gate at time step t , δ cam represents the angle between the optical axis of the camera and the centre of the next gate and \({{\bf{a}}}_{t}^{\omega }\) are the commanded body rates. The hyperparameters λ 1 ,…,  λ 5 balance different terms (Extended Data Table 1a ).

Training details

Data collection is performed by simulating 100 agents in parallel that interact with the environment in episodes of 1,500 steps. At each environment reset, every agent is initialized at a random gate on the track, with bounded perturbation around a state previously observed when passing this gate. In contrast to previous work 44 , 49 , 50 , we do not perform randomization of the platform dynamics at training time. Instead, we perform fine-tuning based on real-world data. The training environment is implemented using TensorFlow Agents 51 . The policy network and the value network are both represented by two-layer perceptrons with 128 nodes in each layer and LeakyReLU activations with a negative slope of 0.2. Network parameters are optimized using the Adam optimizer with learning rate 3 × 10 −4 for both the policy network and the value network.

Policies are trained for a total of 1 × 10 8 environment interactions, which takes 50 min on a workstation (i9 12900K, RTX 3090, 32 GB RAM DDR5). Fine-tuning is performed for 2 × 10 7 environment interactions.

Residual model identification

We perform fine-tuning of the original policy based on a small amount of data collected in the real world. Specifically, we collect three full rollouts in the real world, corresponding to approximately 50 s of flight time. We fine-tune the policy by identifying residual observations and residual dynamics, which are then used for training in simulation. During this fine-tuning phase, only the weights of the control policy are updated, whereas the weights of the gate-detection network are kept constant.

Residual observation model

Navigating at high speeds results in substantial motion blur, which can lead to a loss of tracked visual features and severe drift in linear odometry estimates. We fine-tune policies with an odometry model that is identified from only a handful of trials recorded in the real world. To model the drift in odometry, we use Gaussian processes 36 , as they allow fitting a posterior distribution of odometry perturbations, from which we can sample temporally consistent realizations.

Specifically, the Gaussian process model fits residual position, velocity and attitude as a function of the ground-truth robot state. The observation residuals are identified by comparing the observed visual–inertial odometry (VIO) estimates during a real-world rollout with the ground-truth platform states, which are obtained from an external motion-tracking system.

We treat each dimension of the observation separately, effectively fitting a set of nine 1D Gaussian processes to the observation residuals. We use a mixture of radial basis function kernels

in which L  is the diagonal length scale matrix and σ f and σ n represent the data and prior noise variance, respectively, and z i and z j represent data features. The kernel hyperparameters are optimized by maximizing the log marginal likelihood. After kernel hyperparameter optimization, we sample new realizations from the posterior distribution that are then used during fine-tuning of the policy. Extended Data Fig. 1 illustrates the residual observations in position, velocity and attitude in real-world rollouts, as well as 100 sampled realizations from the Gaussian process model.

Residual dynamics model

We use a residual model to complement the simulated robot dynamics 52 . Specifically, we identify residual accelerations as a function of the platform state s and the commanded mass-normalized collective thrust c :

We use k -nearest neighbour regression with k  = 5. The size of the dataset used for residual dynamics model identification depends on the track layout and ranges between 800 and 1,000 samples for the track layout used in this work.

Gate detection

To correct for drift accumulated by the VIO pipeline, the gates are used as distinct landmarks for relative localization. Specifically, gates are detected in the onboard camera view by segmenting gate corners 26 . The greyscale images provided by the Intel RealSense Tracking Camera T265 are used as input images for the gate detector. The architecture of the segmentation network is a six-level U-Net 53 with (8, 16, 16, 16, 16, 16) convolutional filters of size (3, 3, 3, 5, 7, 7) per level and a final extra layer operating on the output of the U-Net containing 12 filters. As the activation function, LeakyReLU with α  = 0.01 is used. For deployment on the NVIDIA Jetson TX2, the network is ported to TensorRT. To optimize memory footprint and computation time, inference is performed in half-precision mode (FP16) and images are downsampled to size 384 × 384 before being fed to the network. One forward pass through the network takes 40 ms on the NVIDIA Jetson TX2.

VIO drift estimation

The odometry estimates from the VIO pipeline 54 exhibit substantial drift during high-speed flight. We use gate detection to stabilize the pose estimates produced by VIO. The gate detector outputs the coordinates of the corners of all visible gates. A relative pose is first estimated for all predicted gates using infinitesimal plane-based pose estimation (IPPE) 34 . Given this relative pose estimate, each gate observation is assigned to the closest gate in the known track layout, thus yielding a pose estimate for the drone.

Owing to the low frequency of the gate detections and the high quality of the VIO orientation estimate, we only refine the translational components of the VIO measurements. We estimate and correct for the drift of the VIO pipeline using a Kalman filter that estimates the translational drift p d (position offset) and its derivative, the drift velocity v d . The drift correction is performed by subtracting the estimated drift states p d and v d from the corresponding VIO estimates. The Kalman filter state x is given by \({\bf{x}}={[{{\bf{p}}}_{{\rm{d}}}^{\top },{{\bf{v}}}_{{\rm{d}}}^{\top }]}^{\top }\in {{\mathbb{R}}}^{6}\) .

The state x and covariance P  updates are given by:

On the basis of measurements, the process noise is set to σ pos  = 0.05 and σ vel  = 0.1. The filter state and covariance are initialized to zero. For each measurement z k (pose estimate from a gate detection), the predicted VIO drift \({{\bf{x}}}_{k}^{-}\) is corrected to the estimate \({{\bf{x}}}_{k}^{+}\) according to the Kalman filter equations:

in which K k  is the Kalman gain, R  is the measurement covariance and H k is the measurement matrix. If several gates have been detected in a single camera frame, all relative pose estimates are stacked and processed in the same Kalman filter update step. The main source of measurement error is the uncertainty in the gate-corner detection of the network. This error in the image plane results in a pose error when IPPE is applied. We opted for a sampling-based approach to estimate the pose error from the known average gate-corner-detection uncertainty. For each gate, the IPPE algorithm is applied to the nominal gate observation as well as to 20 perturbed gate-corner estimates. The resulting distribution of pose estimates is then used to approximate the measurement covariance R  of the gate observation.

Simulation results

Reaching champion-level performance in autonomous drone racing requires overcoming two challenges: imperfect perception and incomplete models of the system’s dynamics. In controlled experiments in simulation, we assess the robustness of our approach to both of these challenges. To this end, we evaluate performance in a racing task when deployed in four different settings. In setting (1), we simulate a simplistic quadrotor model with access to ground-truth state observations. In setting (2), we replace the ground-truth state observations with noisy observations identified from real-world flights. These noisy observations are generated by sampling one realization from the residual observation model and are independent of the perception awareness of the deployed controller. Settings (3) and (4) share the observation models with the previous two settings, respectively, but replace the simplistic dynamics model with more accurate aerodynamical simulation 43 . These four settings allow controlled assessment of the sensitivity of the approach to changes in the dynamics and the observation fidelity.

In all four settings, we benchmark our approach against the following baselines: zero-shot, domain randomization and time-optimal. The zero-shot baseline represents a learning-based racing policy 35 trained using model-free RL that is deployed zero-shot from the training domain to the test domain. The training domain of the policy is equal to experimental setting (1), that is, idealized dynamics and ground-truth observations. Domain randomization extends the learning strategy from the zero-shot baseline by randomizing observations and dynamics properties to increase robustness. The time-optimal baseline uses a precomputed time-optimal trajectory 28 that is tracked using an MPC controller. This approach has shown the best performance in comparison with other model-based methods for time-optimal flight 55 , 56 . The dynamics model used by the trajectory generation and the MPC controller matches the simulated dynamics of experimental setting (1).

Performance is assessed by evaluating the fastest lap time, the average and minimum observed gate margin of successfully passed gates and the percentage of track successfully completed. The gate margin metric measures the distance between the drone and the closest point on the gate when crossing the gate plane. A high gate margin indicates that the quadrotor passed close to the centre of the gate. Leaving a smaller gate margin can increase speed but can also increase the risk of collision or missing the gate. Any lap that results in a crash is not considered valid.

The results are summarized in Extended Data Table 1c . All approaches manage to successfully complete the task when deployed in idealized dynamics and ground-truth observations, with the time-optimal baseline yielding the lowest lap time. When deployed in settings that feature domain shift, either in the dynamics or the observations, the performance of all baselines collapses and none of the three baselines are able to complete even a single lap. This performance drop is exhibited by both learning-based and traditional approaches. By contrast, our approach, which features empirical models of dynamics and observation noise, succeeds in all deployment settings, with small increases in lap time.

The key feature that enables our approach to succeed across deployment regimes is the use of an empirical model of dynamics and observation noise, estimated from real-world data. A comparison between an approach that has access to such data and approaches that do not is not entirely fair. For that reason, we also benchmark the performance of all baseline approaches when having access to the same real-world data used by our approach. Specifically, we compare the performance in experimental setting (2), which features the idealized dynamics model but noisy perception. All baseline approaches are provided with the predictions of the same Gaussian process model that we use to characterize observation noise. The results are summarized in Extended Data Table 1b . All baselines benefit from the more realistic observations, yielding higher completion rates. Nevertheless, our approach is the only one that reliably completes the entire track. As well as the predictions of the observation noise model, our approach also takes into account the uncertainty of the model. For an in-depth comparison of the performance of RL versus optimal control in controlled experiments, we refer the reader to ref.  57 .

Fine-tuning for several iterations

We investigate the extent of variations in behaviour across iterations. The findings of our analysis reveal that subsequent fine-tuning operations result in negligible enhancements in performance and alterations in behaviour (Extended Data Fig. 2 ).

In the following, we provide more details on this investigation. We start by enumerating the fine-tuning steps to provide the necessary notation:

Train policy-0 in simulation.

Deploy policy-0 in the real world. The policy operates on ground-truth data from a motion-capture system.

Identify residuals observed by policy-0 in the real world.

Train policy-1 by fine-tuning policy-0 on the identified residuals.

Deploy policy-1 in the real world. The policy operates only on onboard sensory measurements.

Identify residuals observed by policy-1 in the real world.

Train policy-2 by fine-tuning policy-1 on the identified residuals.

We compare the performance of policy-1 and policy-2 in simulation after fine-tuning on their respective residuals. The results are illustrated in Extended Data Fig. 2 . We observe that the difference in distance from gate centres, which is a metric for the safety of the policy, is 0.09 ± 0.08 m. Furthermore, the difference in the time taken to complete a single lap is 0.02 ± 0.02 s. Note that this lap-time difference is substantially smaller than the difference between the single-lap completion times of Swift and the human pilots (0.16 s).

Drone hardware configuration

The quadrotors used by the human pilots and Swift have the same weight, shape and propulsion. The platform design is based on the Agilicious framework 58 . Each vehicle has a weight of 870 g and can produce a maximum static thrust of approximately 35 N, which results in a static thrust-to-weight ratio of 4.1. The base of each platform consists of an Armattan Chameleon 6″ main frame that is equipped with T-Motor Velox 2306 motors and 5″, three-bladed propellers. An NVIDIA Jetson TX2 accompanied by a Connect Tech Quasar carrier board provides the main compute resource for the autonomous drones, featuring a six-core CPU running at 2 GHz and a dedicated GPU with 256 CUDA cores running at 1.3 GHz. Although forward passes of the gate-detection network are performed on the GPU, the racing policy is evaluated on the CPU, with one inference pass taking 8 ms. The autonomous drones carry an Intel RealSense Tracking Camera T265 that provides VIO estimates 59 at 100 Hz that are fed by USB to the NVIDIA Jetson TX2. The human-piloted drones carry neither a Jetson computer nor a RealSense camera and are instead equipped with a corresponding ballast weight. Control commands in the form of collective thrust and body rates produced by the human pilots or Swift are sent to a commercial flight controller, which runs on an STM32 processor operating at 216 MHz. The flight controller is running Betaflight, an open-source flight-control software 45 .

Human pilot impressions

The following quotes convey the impressions of the three human champions who raced against Swift.

Alex Vanover :

These races will be decided at the split S, it is the most challenging part of the track.

This was the best race! I was so close to the autonomous drone, I could really feel the turbulence when trying to keep up with it.

Thomas Bitmatta :

The possibilities are endless, this is the start of something that could change the whole world. On the flip side, I’m a racer, I don’t want anything to be faster than me.

As you fly faster, you trade off precision for speed.

It’s inspiring to see the potential of what drones are actually capable of. Soon, the AI drone could even be used as a training tool to understand what would be possible.

Marvin Schaepper :

It feels different racing against a machine, because you know that the machine doesn’t get tired.

Research ethics

The study has been conducted in accordance with the Declaration of Helsinki. The study protocol is exempt from review by an ethics committee according to the rules and regulations of the University of Zurich, because no health-related data has been collected. The participants gave their written informed consent before participating in the study.

Data availability

All (other) data needed to evaluate the conclusions in the paper are present in the paper or the extended data. Motion-capture recordings of the race events with accompanying analysis code can be found in the file ‘racing_data.zip’ on Zenodo at https://doi.org/10.5281/zenodo.7955278 .

Code availability

Pseudocode for Swift detailing the training process and algorithms can be found in the file ‘pseudocode.zip’ on Zenodo at https://doi.org/10.5281/zenodo.7955278 . To safeguard against potential misuse, the full source code associated with this research will not be made publicly available.

De Wagter, C., Paredes-Vallés, F., Sheth, N. & de Croon, G. Learning fast in autonomous drone racing. Nat. Mach. Intell. 3 , 923 (2021).

Article   Google Scholar  

Hanover, D. et al. Autonomous drone racing: a survey. Preprint at https://arxiv.org/abs/2301.01755 (2023).

Sutton, R. S. & Barto, A. G. Reinforcement Learning: An Introduction (MIT Press, 2018).

Mnih, V. et al. Human-level control through deep reinforcement learning. Nature 518 , 529–533 (2015).

Article   ADS   CAS   PubMed   Google Scholar  

Schrittwieser, J. et al. Mastering Atari, Go, chess and shogi by planning with a learned model. Nature 588 , 604–609 (2020).

Ecoffet, A., Huizinga, J., Lehman, J., Stanley, K. O. & Clune, J. First return, then explore. Nature 590 , 580–586 (2021).

Silver, D. et al. Mastering the game of Go with deep neural networks and tree search. Nature 529 , 484–489 (2016).

Silver, D. et al. Mastering the game of Go without human knowledge. Nature 550 , 354–359 (2017).

Silver, D. et al. A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science 362 , 1140–1144 (2018).

Article   ADS   MathSciNet   CAS   PubMed   MATH   Google Scholar  

Vinyals, O. et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature 575 , 350–354 (2019).

Berner, C. et al. Dota 2 with large scale deep reinforcement learning. Preprint at https://arxiv.org/abs/1912.06680 (2019).

Fuchs, F., Song, Y., Kaufmann, E., Scaramuzza, D. & Dürr, P. Super-human performance in Gran Turismo Sport using deep reinforcement learning. IEEE Robot. Autom. Lett. 6 , 4257–4264 (2021).

Wurman, P. R. et al. Outracing champion Gran Turismo drivers with deep reinforcement learning. Nature 602 , 223–228 (2022).

Funke, J. et al. in Proc. 2012 IEEE Intelligent Vehicles Symposium 541–547 (IEEE, 2012).

Spielberg, N. A., Brown, M., Kapania, N. R., Kegelman, J. C. & Gerdes, J. C. Neural network vehicle models for high-performance automated driving. Sci. Robot. 4 , eaaw1975 (2019).

Article   PubMed   Google Scholar  

Won, D.-O., Müller, K.-R. & Lee, S.-W. An adaptive deep reinforcement learning framework enables curling robots with human-like performance in real-world conditions. Sci. Robot. 5 , eabb9764 (2020).

Moon, H., Sun, Y., Baltes, J. & Kim, S. J. The IROS 2016 competitions. IEEE Robot. Autom. Mag. 24 , 20–29 (2017).

Jung, S., Hwang, S., Shin, H. & Shim, D. H. Perception, guidance, and navigation for indoor autonomous drone racing using deep learning. IEEE Robot. Autom. Lett. 3 , 2539–2544 (2018).

Kaufmann, E. et al. in Proc. 2nd Conference on Robot Learning (CoRL) 133–145 (PMLR, 2018).

Zhang, D. & Doyle, D. D. in Proc. 2020 IEEE Aerospace Conference , 1–11 (IEEE, 2020).

Loquercio, A. et al. Deep drone racing: from simulation to reality with domain randomization. IEEE Trans. Robot. 36 , 1–14 (2019).

Loquercio, A. et al. Learning high-speed flight in the wild. Sci. Robot. 6 , eabg5810 (2021).

Kaufmann, E. et al. in Proc. 2019 International Conference on Robotics and Automation (ICRA) 690–696 (IEEE, 2019).

Li, S., van der Horst, E., Duernay, P., De Wagter, C. & de Croon, G. C. Visual model-predictive localization for computationally efficient autonomous racing of a 72-g drone. J. Field Robot. 37 , 667–692 (2020).

A.I. is flying drones (very, very slowly). https://www.nytimes.com/2019/03/26/technology/alphapilot-ai-drone-racing.html (2019).

Foehn, P. et al. AlphaPilot: autonomous drone racing. Auton. Robots 46 , 307–320 (2021).

Article   PubMed   PubMed Central   Google Scholar  

Wagter, C. D., Paredes-Vallé, F., Sheth, N. & de Croon, G. The sensing, state-estimation, and control behind the winning entry to the 2019 Artificial Intelligence Robotic Racing Competition. Field Robot. 2 , 1263–1290 (2022).

Foehn, P., Romero, A. & Scaramuzza, D. Time-optimal planning for quadrotor waypoint flight. Sci. Robot. 6 , eabh1221 (2021).

Romero, A., Sun, S., Foehn, P. & Scaramuzza, D. Model predictive contouring control for time-optimal quadrotor flight. IEEE Trans. Robot. 38 , 3340–3356 (2022).

Sun, S., Romero, A., Foehn, P., Kaufmann, E. & Scaramuzza, D. A comparative study of nonlinear MPC and differential-flatness-based control for quadrotor agile flight. IEEE Trans. Robot. 38 , 3357–3373 (2021).

Schulman, J., Wolski, F., Dhariwal, P., Radford, A. & Klimov, O. Proximal policy optimization algorithms. Preprint at https://arxiv.org/abs/1707.06347 (2017).

Scaramuzza, D. & Zhang, Z. Encyclopedia of Robotics (eds Ang, M., Khatib, O. & Siciliano, B.) 1–9 (Springer, 2019).

Huang, G. in Proc. 2019 International Conference on Robotics and Automation (ICRA) 9572–9582 (IEEE, 2019).

Collins, T. & Bartoli, A. Infinitesimal plane-based pose estimation. Int. J. Comput. Vis. 109 , 252–286 (2014).

Article   MathSciNet   MATH   Google Scholar  

Song, Y., Steinweg, M., Kaufmann, E. & Scaramuzza, D. in Proc. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 1205–1212 (IEEE, 2021).

Williams, C. K. & Rasmussen, C. E. Gaussian Processes for Machine Learning (MIT Press, 2006).

Hwangbo, J. et al. Learning agile and dynamic motor skills for legged robots. Sci. Robot. 4 , eaau5872 (2019).

Hung, C.-C. et al. Optimizing agent behavior over long time scales by transporting value. Nat. Commun. 10 , 5223 (2019).

Article   ADS   PubMed   PubMed Central   Google Scholar  

Pfeiffer, C. & Scaramuzza, D. Human-piloted drone racing: visual processing and control. IEEE Robot. Autom. Lett. 6 , 3467–3474 (2021).

Spica, R., Cristofalo, E., Wang, Z., Montijano, E. & Schwager, M. A real-time game theoretic planner for autonomous two-player drone racing. IEEE Trans. Robot. 36 , 1389–1403 (2020).

Day, B. L. & Fitzpatrick, R. C. The vestibular system. Curr. Biol. 15 , R583–R586 (2005).

Article   CAS   PubMed   Google Scholar  

Kim, J. et al. Esports arms race: latency and refresh rate for competitive gaming tasks. J. Vis. 19 , 218c (2019).

Bauersfeld, L., Kaufmann, E., Foehn, P., Sun, S. & Scaramuzza, D. in Proc. Robotics: Science and Systems XVII 42 (Robotics: Science and Systems Foundation, 2021).

Kaufmann, E., Bauersfeld, L. & Scaramuzza, D. in Proc. 2022 International Conference on Robotics and Automation (ICRA) 10504–10510 (IEEE, 2022).

The Betaflight Open Source Flight Controller Firmware Project. Betaflight. https://github.com/betaflight/betaflight (2022).

Bauersfeld, L. & Scaramuzza, D. Range, endurance, and optimal speed estimates for multicopters. IEEE Robot. Autom. Lett. 7 , 2953–2960 (2022).

Zhou, Y., Barnes, C., Lu, J., Yang, J. & Li, H. in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 5745–5753 (IEEE, 2019).

Pinto, L., Andrychowicz, M., Welinder, P., Zaremba, W. & Abbeel, P. in Proc. Robotics: Science and Systems XIV (MIT Press Journals, 2018).

Molchanov, A. et al. in Proc. 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 59–66 (IEEE, 2019).

Andrychowicz, O. M. et al. Learning dexterous in-hand manipulation. Int. J. Robot. Res. 39 , 3–20 (2020).

Guadarrama, S. et al. TF-Agents: a library for reinforcement learning in TensorFlow. https://github.com/tensorflow/agents (2018).

Torrente, G., Kaufmann, E., Foehn, P. & Scaramuzza, D. Data-driven MPC for quadrotors. IEEE Robot. Autom. Lett. 6 , 3769–3776 (2021).

Ronneberger, O., Fischer, P. & Brox, T. in Proc. International Conference on Medical Image Computing and Computer-assisted Intervention 234–241 (Springer, 2015).

Intel RealSense T265 series product family. https://www.intelrealsense.com/wp-content/uploads/2019/09/Intel_RealSense_Tracking_Camera_Datasheet_Rev004_release.pdf (2019).

Ryou, G., Tal, E. & Karaman, S. Multi-fidelity black-box optimization for time-optimal quadrotor maneuvers. Int. J. Robot. Res. 40 , 1352–1369 (2021).

Pham, H. & Pham, Q.-C. A new approach to time-optimal path parameterization based on reachability analysis. IEEE Trans. Robot. 34 , 645–659 (2018).

Song, Y., Romero, A., Müller, M., Koltun, V. & Scaramuzza, D. Reaching the limit in autonomous racing: optimal control versus reinforcement learning. Sci. Robot. (in the press).

Foehn, P. et al. Agilicious: open-source and open-hardware agile quadrotor for vision-based flight. Sci. Robot. 7 , eabl6259 (2022).

Jones, E. S. & Soatto, S. Visual-inertial navigation, mapping and localization: a scalable real-time causal approach. Int. J. Robot. Res. 30 , 407–430 (2011).

Download references

Acknowledgements

The authors thank A. Vanover, T. Bitmatta and M. Schaepper for accepting to race against Swift. The authors also thank C. Pfeiffer, T. Längle and A. Barden for their contributions to the organization of the race events and the drone hardware design. This work was supported by Intel’s Embodied AI Lab, the Swiss National Science Foundation (SNSF) through the National Centre of Competence in Research (NCCR) Robotics and the European Research Council (ERC) under grant agreement 864042 (AGILEFLIGHT).

Author information

Authors and affiliations.

Robotics and Perception Group, University of Zurich, Zürich, Switzerland

Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio & Davide Scaramuzza

Intel Labs, Munich, Germany

Matthias Müller

Intel Labs, Jackson, WY, USA

Vladlen Koltun

You can also search for this author in PubMed   Google Scholar

Contributions

E.K. formulated the main ideas, implemented the system, performed the experiments and data analysis and wrote the paper. L.B. contributed to the main ideas, the experiments, data analysis, paper writing and designed the graphical illustrations. A.L. formulated the main ideas and contributed to the experimental design, data analysis and paper writing. M.M. contributed to the experimental design, data analysis and paper writing. V.K. contributed to the main ideas, the experimental design, the analysis of experiments and paper writing. D.S. contributed to the main ideas, experimental design, analysis of experiments, paper writing and provided funding.

Corresponding author

Correspondence to Elia Kaufmann .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Peer review

Peer review information.

Nature thanks Sunggoo Jung, Karime Pereida and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended data fig. 1 residual models..

a , Visualization of the residual observation model and the residual dynamics model identified from real-world data. Black curves depict the residual observed in the real world and coloured lines show 100 sampled realizations of the residual observation model. Each plot depicts an entire race, that is, three laps. b , Predicted residual observation for a simulated rollout. Blue, ground-truth position provided by the simulator; orange, perturbed position generated by the Gaussian process residual.

Extended Data Fig. 2 Multi-iteration fine-tuning.

Rollout comparison after fine-tuning the policy for one iteration (blue) and two iterations (orange).

Supplementary information

Peer review file, rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and Permissions

About this article

Cite this article.

Kaufmann, E., Bauersfeld, L., Loquercio, A. et al. Champion-level drone racing using deep reinforcement learning. Nature 620 , 982–987 (2023). https://doi.org/10.1038/s41586-023-06419-4

Download citation

Received : 05 January 2023

Accepted : 10 July 2023

Published : 30 August 2023

Issue Date : 31 August 2023

DOI : https://doi.org/10.1038/s41586-023-06419-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Read our research on: World Leaders | Artificial Intelligence | Science Issues

Regions & Countries

Most americans haven’t used chatgpt; few think it will have a major impact on their job.

The debut of ChatGPT has led some tech experts to declare it part of a “robot revolution.” But most Americans haven’t used ChatGPT, and only a small share think chatbots will have a major impact on their jobs. Even fewer Americans say chatbots would be helpful for their own work, according to a new Pew Research Center survey conducted July 17-23.

Pew Research Center has long explored Americans’ perspectives on emerging technologies and uses of artificial intelligence. The current study examined adults’ experiences with chatbots such as ChatGPT, particularly in relation to their jobs. This survey was conducted among 5,057 U.S. adults from July 17 to 23, 2023. Everyone who took part in the survey is a member of the Center’s American Trends Panel (ATP), an online survey panel that is recruited through national, random sampling of residential addresses. This way nearly all U.S. adults have a chance of selection. The survey is weighted to be representative of the U.S. adult population by gender, race and ethnicity, partisan affiliation, education and other categories. Read more about the  ATP’s methodology .

Here are the questions used for this analysis , along with responses, and its methodology .

Who has used ChatGPT?

Among those who have heard of ChatGPT, 24% say they have ever used it. That amounts to 18% of U.S. adults overall.

• Younger adults are more likely than older adults to have used ChatGPT. Among those who have heard about ChatGPT, about four-in-ten adults under 30 have used it; that share drops to 19% when looking at 50- to 64-year-olds and to just 5% for those 65 and older.
• College-educated Americans stand out: 32% of adults who have a college degree or more education and have heard of ChatGPT have used it.
• Among those who have heard of the chatbot, men are more likely than women to report using ChatGPT (29% vs. 19%).

How do Americans use ChatGPT?

One-in-five U.S. adults who have heard of ChatGPT say they have ever used it for entertainment. A similar share (19%) say they have used it to learn something, while 16% of those who have heard of the tool and are employed say they have used it for tasks at work.

Younger adults are more likely than their older peers to have used ChatGPT for education or amusement. For example, among those who have heard of ChatGPT, three-in-ten adults under 30 have used it for learning, compared with 11% of those 50 and older.

Similarly, 29% of adults who are under 50 and have heard of ChatGPT have used it for entertainment, while 10% of their older counterparts have done the same.

There is a similar pattern among employed adults who have heard of the tool: Those under 50 are more likely than those 50 and older to report having used ChatGPT for work (18% vs. 10%).

Adults with a high school education or less who have heard of ChatGPT are significantly less likely than their peers with higher levels of formal education to have used ChatGPT for entertainment, learning or work. For example, among those who have heard of ChatGPT, 13% who have a high school diploma or less say they have used it to learn something new, compared with roughly one-in-five or more of those with some college (21%), a bachelor’s degree (19%) or a postgraduate education (26%).

How has ChatGPT use changed recently?

More Americans who have heard of ChatGPT are using it for work and learning today than when the Center first asked about this in the spring . For example, the share of employed adults who have heard of ChatGPT and have used it for work tasks increased from 12% in March to 16% in the new survey, which was conducted in July. The share who have used ChatGPT to learn something has also risen slightly during this time.

Which industries do Americans think will be most affected by chatbots?

Generative artificial intelligence like ChatGPT may have its greatest impact on jobs that were traditionally thought to be immune from automation – namely, higher-paying jobs that require a college education .

In the Center’s new survey, about half or more of those who have heard of ChatGPT say chatbots will have a major impact on software engineers (56%), graphic designers (54%) and journalists (52%) over the next 20 years. Smaller shares think chatbots will have a major effect on teachers (44%) or lawyers (31%).  

But Americans are less likely to think chatbots will impact their own job. Some 19% of employed adults who have heard of ChatGPT think chatbots will have a major impact on their job. Another 36% say it will have a minor impact and 27% expect no impact at all.

Which Americans are most likely to think chatbots will affect their own jobs?

Among employed adults who have heard of ChatGPT, 60% of those ages 30 to 49 say chatbots will have a major or minor impact on their own job in the next 20 years. This is the highest percentage of any age group.

Those with more formal education are also more likely to think chatbots will have an impact on their job. For example, among working adults who have heard of ChatGPT, 71% of those who have a postgraduate degree say this, compared with 60% of those with a bachelor’s degree only and even smaller shares of those with some college (48%) or a high school diploma or less (42%).

Views on this question also vary by industry.

Those who work in information and technology and have heard of ChatGPT are the most likely to say chatbots will have a major or minor impact on their job. Three-quarters of these workers say this, including 37% who say chatbots will have a major impact on their work.

Around two-thirds (66%) of those who have heard of ChatGPT and work in education banking, finance, accounting, real estate or insurance think chatbots will affect their job.

At the other end of the spectrum, 40% of adults who have heard of ChatGPT and work in hospitality, service, arts, entertainment or recreation say chatbots will affect their job.

Will chatbots help people do their jobs?

Most employed Americans don’t anticipate chatbots being very helpful for their own work. Only 15% of working adults who have heard of ChatGPT say chatbots would be extremely or very helpful for their job, with 5% saying they would be extremely helpful.

Younger workers who have heard of ChatGPT are more optimistic that chatbots would help them do their job.

Workers with higher levels of formal education are also more optimistic. Those who have heard of ChatGPT and have a college degree are slightly more likely than those without a college degree to say these types of tools would be helpful for their job.

How do Americans feel about government regulation of chatbots?

The emergence of ChatGPT has sparked conversations about government regulation of artificial intelligence.  

Asked which is their greater concern, two-thirds of those who have heard of ChatGPT say it’s that government will not go far enough in regulating chatbot use. Some 31% instead say their greater concern is that the government will go too far.  

Of those who have heard of ChatGPT, majorities of Democrats and Republicans say their greater concern is not enough regulation. But this view is more common among Democrats and Democratic-leaning independents than among their Republican and Republican-leaning counterparts (75% vs. 59%).

Note: Here are the questions used for this analysis , along with responses, and its methodology .

Sign up for our weekly newsletter

Fresh data delivered Saturday mornings

Growing public concern about the role of artificial intelligence in daily life

Which u.s. workers are more exposed to ai on their jobs, as ai spreads, experts predict the best and worst changes in digital life by 2035, a majority of americans have heard of chatgpt, but few have tried it themselves, most popular.

About Pew Research Center Pew Research Center is a nonpartisan fact tank that informs the public about the issues, attitudes and trends shaping the world. It conducts public opinion polling, demographic research, media content analysis and other empirical social science research. Pew Research Center does not take policy positions. It is a subsidiary of The Pew Charitable Trusts .

Dailymotion

Research shows systemic racism still exists in job market

People with non-English names get almost half the call backs for job applications compared to those with English names. Research studies continue to find systemic racism still exists in the job market with qualified applicants often missing out because of their race or ethnicity. And some young people are resorting to altering their identity on paper to get a fair go.

bottom_desktop desktop:[300x250]

Not so eco-friendly? Paper straws contain more 'forever chemicals' than plastic, study says

Not a fan of those paper straws that have replaced disposable plastic ones in the name of being eco-friendly? As it turns out, some of those efforts to save the environment may have been in vain.

A new study, published Thursday in the journal Food Additives and Contaminants, found evidence of “forever chemical” PFAS (per- and polyfluoroalkyl substances) in the majority of both paper and bamboo straws tested.

Scientists in Belgium tested 39 brands of straws made of paper, bamboo, plastic and stainless steel found in shops, supermarkets and restaurants across the country.

Of the straws tested, almost all contained some concertation of PFAS, which are often used during manufacturing to make products water resistant. Out of the total 39 tested, the chemicals were detected in 27, none of which were stainless steel.

Paper straws, on the other hand, were the most likely to contain PFAS, with 18 out of 20, or 90%, of paper brands testing positive. They were also found in four out of five bamboo straws, three out of four plastic straws and two out of five glass straws.

Not so eco-friendly?

Eighteen different PFAS were detected in total, though overall in low concentrations. The chemical most commonly found, however, was perfluorooctanoic acid (PFOA) which was banned globally in 2020.

“These ‘eco-friendly’ plant-based straws are not necessarily a more sustainable alternative to plastic straws,” said the study’s conclusion, “because they can be considered as an additional source of PFAS exposure in humans and the environment (e.g. after degradation in landfills or through incomplete incineration).”

The study also discovered PFAS that are known to be highly water soluble, meaning they have the potential to bleed from the straw into a drink, but did not investigate this component further.

The researchers proposed that, while manufactures could intentionally be coating their plant-based straws in chemicals to make them water-repellent, the presence of PFAS could also be attributed to contaminated soil or an unintended consequence of material recycling. The authors suggested further analysis and studies be conducted to determine the primary source of contamination in the straws and how the chemicals may impact drinks and people consuming them.

This Belgian study comes on the heels of a 2021 U.S. study, which found the presence of 21 PFAS in paper and other plant-based straws versus no measurable amounts in plastic ones.

While PFAS were present in most straws tested, the low concentration, paired with the limited extent to which people use straws, means they don’t pose an immediate risk to humans.

Small amounts of PFAS are not harmful in and of themselves, but rather their ability to build up over time, including in the human body, is what poses the most risk. Even with these findings, plant-based straws are still better for the environment than straight-up plastics.

As stainless-steel straws are reusable long-term and all tested PFAS-free, the study authors suggest the use of these straws for bother environmental and health-related reasons.

PFAS found in drinking water: Dangerous levels of PFAS detected in water for 27 million. Did the EPA find it near you?

What are PFAS?

PFAS stands for “per- and polyfluoroalkyl substances,” and refers to a collection of long-lasting chemicals that take a very long time to slowly break down in the environment.

According to the United States Environmental Protection Agency (EPA), PFAS are widely used and persist for long periods of time in the environment, meaning they are found in the blood of people and animals around the world, as well as air, water, soil and in low levels in foods, packaging and household products.

What health risks are associated with PFAS?

While scientists are still working to determine the extent to which PFAS impact us, animals and our environment, they are already associated with a list of health concerns.

According to the EPA, PFAS have been linked to:

• Reproductive effects such as decreased fertility or increased high blood pressure in pregnant women.
• Developmental effects or delays in children, including low birth weight, accelerated puberty, bone variations, or behavioral changes.
• Increased risk of some cancers, including prostate, kidney, and testicular cancers.
• Reduced ability of the body’s immune system to fight infections, including reduced vaccine response.
• Interference with the body’s natural hormones.
• Increased cholesterol levels and/or risk of obesity.

Eye drop recall: 2 more eye drop products recalled after being linked to potentially deadly bacteria

Where are PFAS usually found?

According to the U.S. Food and Drug Administration (FDA), PFAS can be found in hundreds of products we use daily. In some cases, they are approved for use in limited amounts by the FDA, such as in food packaging.

They are commonly found in:

• Stain- and water-resistant fabrics and carpeting.
• Cleaning products.
• Fire-fighting foams.
• Food packaging.
• Food processing equipment.

Adani family partners used offshore funds to invest in Indian group's stocks, report says

The logo of the Adani Group is seen on the facade of its Corporate House on the outskirts of Ahmedabad, India, January 27, 2023. REUTERS/Amit Dave/File Photo Acquire Licensing Rights

• Adani Enterprises Ltd Follow
• Adani Group Follow
• 360 One Wam Ltd Follow

NEW DELHI, Aug 31 (Reuters) - Millions of dollars were invested in publicly traded Adani Group stocks through funds in Mauritius, the Organised Crime and Corruption Reporting Project (OCCRP) reported on Thursday, saying this "obscured" the involvement of alleged business partners of India's Adani family.

Citing a review of files from tax havens and internal Adani Group emails, the non-profit global network of investigative journalists said two individual investors with "longtime business ties" to the Adani family used such offshore structures to buy and sell Adani shares between 2013 and 2018.

The Adani Group, which is controlled by billionaire Gautam Adani, said it categorically rejected what it called recycled allegations in the OCCRP report "in their entirety".

Nasser Ali Shaban Ahli from Dubai and Chang Chung-Ling from Taiwan, the two investors named in the OCCRP report, did not respond to Reuters requests seeking comment.

Reuters has not independently verified the allegations made in the OCCRP report, which comes after U.S.-based short-seller Hindenburg Research accused the Adani Group in January of improper business dealings .

Shares in Adani Group companies fell on Thursday amid renewed corporate governance concerns. Adani Enterprises (ADEL.NS) , the group's flagship company, closed down 3.7%, while Adani Ports (APSE.NS) , Adani Power (ADAN.NS) , Adani Green (ADNA.NS) , Adani Total Gas (ADAG.NS) , Adani Energy Solutions (ADAI.NS) and Adani Wilmar (ADAW.NS) slid between 2% and 4.3%.

"If true, it could mean a violation of Indian financial market regulator SEBI laws for publicly listed stocks, that could sway the outcome or push SEBI to dig deeper in its ongoing investigation into the group," CreditiSights senior research analyst Lakshmanan R. said.

The Securities and Exchange Board of India (SEBI), did not officially respond to Reuters' requests for comment.

Sources told Reuters that SEBI has examined the two Mauritius-based funds and one Bermuda-based fund cited by OCCRP as part of the regulator's larger probe into the Adani Group.

The investigation into possible violation of public float norms by the Adani Group is still ongoing and any new facts will be considered, those sources added.

In the days following the January report, Adani Group stocks lost $150 billion in market value and remain down around $100 billion following a recovery in recent months after it repaid some debt and regained some investor confidence.

FAMILY TIES

Between them, at the peak of their investment in June 2016, Ahli and Chang held free-floating shares of four Adani Group units - Adani Power, Adani Enterprises, Adani Ports, and Adani Energy Solutions (formerly known as Adani Transmission) - ranging from 8% to about 14% stakes in the companies through two Mauritius-based funds, the OCCRP report said.

At one point, their investment in Adani funds was worth $430 million, the report said.

Under Indian laws, every company needs to have 25% of its shares held by public shareholders to avoid price manipulation.

While OCCRP said there was no evidence Chang and Ahli's money for their investments came from the Adani family, its reporting and documents - including an agreement, corporate records and an email - showed there "is evidence" that their trading in Adani stock "was coordinated with the family."

It said that Ahli and Chang were associated with companies of the group as well as with Vinod Adani, who is a brother of Gautam Adani. Vinod Adani did not respond to a Reuters request for comment.

"The question of whether this arrangement is a violation of the law rests on whether Ahli and Chang should be considered to be acting on behalf of Adani 'promoters,' a term used in India to refer to the majority owners of a business," OCCRP said.

If so, OCCRP said, the stake of promoters in Adani holdings would exceed the 75% limit allowed for insider ownership.

Indian asset management services provider 360 One Wam (ONEW.NS) , whose Mauritius arm managed the Emerging India Focus Fund and EM Resurgent Fund that were cited by OCCRP, said the funds sold their investments in Adani stocks in 2018.

'LOOP CLOSED'

Hindenburg said on platform X on Thursday that the OCCRP report closed the loop on issues it had flagged with respect to the offshore funds owning at least 13% of the public float in multiple Adani stocks through "associates of Vinod Adani".

Adani Group had called Hindenburg's claims misleading and without evidence and said it always complied with laws.

In a statement to OCCRP, Adani Group said the Mauritius funds investigated by reporters had already been named in the Hindenburg report and the "allegations are not only baseless and unsubstantiated but are rehashed from Hindenburg's allegations".

India's Supreme Court has appointed a panel to oversee a SEBI probe based on the Hindenburg report. The panel in May said the regulator had so far " drawn a blank " in investigations into the suspected violations.

Last week, SEBI said its report was nearing completion and its investigation on some offshore deals was taking time as some entities were located in tax haven jurisdictions. The regulator "shall take appropriate action based on outcome of the investigations," it said.

SEBI also said it examined one Adani group transaction for violation of minimum public float rules, an issue that the OCCRP report also flagged.

In an interview with a reporter from the Guardian, OCCRP said Chang said he knew nothing about any secret purchases of Adani stock. He asked why journalists were not interested in his other investments and said, "We are a simple business."

Meanwhile, India’s main opposition leader Rahul Gandhi repeated demands for a parliamentary probe given the latest allegations, asking Prime Minister Narendra Modi to clear "his name and categorically explain what is going on".

Indian opposition parties allege that Gautam Adani has benefited from what they say are his close ties with Modi for over two decades, a charge rejected by both Modi and Adani.

Reporting by Aditya Kalra, Krishn Kaushik, Scott Murdoch, Sethuraman NR and Jayshree P Upadhyay; Editing by Lisa Shumaker, Muralikumar Anantharaman, Dhanya Skariachan, Raju Gopalakrishnan and Alexander Smith

Our Standards: The Thomson Reuters Trust Principles.

Thomson Reuters

Krishn reports on politics and strategic affairs from the Indian subcontinent. He has previously worked at the Organized Crime and Corruption Reporting Project, an international investigative consortium; The Indian Express; and The Caravan magazine, writing about defence, politics, law, conglomerates, media, elections and investigative projects. A graduate of Columbia University's journalism school, Krishn has won multiple awards for his work. Contact: +918527322283

Aditya Kalra is the Company News Editor for Reuters in India, overseeing business coverage and reporting stories on some of the world's biggest companies. He joined Reuters in 2008 and has in recent years written stories on challenges and strategies of a wide array of companies -- from Amazon, Google and Walmart to Xiaomi, Starbucks and Reliance. He also extensively works on deeply-reported and investigative business stories.

Scott Murdoch has been a journalist for more than two decades working for Thomson Reuters and News Corp in Australia. He has specialised in financial journalism for most of his career and covers equity and debt capital markets across Asia and Australian M&A. He is based in Sydney.

US private funds industry sues securities regulator over new rules

U.S. judge approves payouts from Elon Musk's SEC settlement

RBC's $10 billion deal for HSBC unit approved by Canada's Competition Bureau

Amgen gets FTC's go-ahead for $27.8 billion Horizon deal

More from Reuters

Belarus says Polish helicopter crossed border, Warsaw denies it

Belarus, Russia's closest ally, summoned a Polish diplomat to protest what it said was a Polish military helicopter's violation of its border on Friday, but a military official in Warsaw denied any incursion had occurred.

'The boom is over': Venezuelans lament end of brief dollarization boost

Western creditors and Cuba pledge to salvage debt deal

India cuts windfall tax on petroleum crude

French President Macron: I talk 'every day' to Niger's ousted president Bazoum