mit mini cheetah thesis

Help | Advanced Search

Computer Science > Robotics

Title: real-time optimal landing control of the mit mini cheetah.

Abstract: Quadrupedal landing is a complex process involving large impacts, elaborate contact transitions, and is a crucial recovery behavior observed in many biological animals. This work presents a real-time, optimal landing controller that is free of pre-specified contact schedules. The controller determines optimal touchdown postures and reaction force profiles and is able to recover from a variety of falling configurations. The quadrupedal platform used, the MIT Mini Cheetah, recovered safely from drops of up to 8 m in simulation, as well as from a range of orientations and planar velocities. The controller is also tested on hardware, successfully recovering from drops of up to 2 m.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

DBLP - CS Bibliography

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 .

Mini Cheetah Clone Teardown, By None Other Than Original Designer

[Ben Katz] designed the original MIT Mini Cheetah robot, which easily captured attention and imagination with its decidedly un-robotic movements and backflips. Not long after [Ben]’s masters thesis went online, clones of the actuators started to show up at overseas sellers, and a few months after that, clones of the whole robot. [Ben] recently had the opportunity to disassemble just such a clone by Dogotix and see what was inside.

Amusingly, one of the first things he noticed is that the “feet” are still just off-the-shelf squash balls, same as his original mini cheetah design. As for the rest of the leg, inside is a belt that goes past some tensioners, connecting the knee joint to an actuator in the shoulder.

As one may expect, these parts are subject to a fair bit of stress, so they have to be sturdy. This design allows for slender yet strong legs without putting an actuator in the knee joint, and you may recall we’ve seen a similar robot gain the ability to stand with the addition of a rigid brace .

It’s interesting to read [Ben]’s thoughts as he disassembles and photographs the unit, and you’ll have to read his post to catch them all. But in the meantime, why not take a moment to see how a neighbor’s curious sheep react to the robot in the video embedded below? The robot botches a backflip due to a low battery, but the sheep seem suitably impressed anyway.

21 thoughts on “ Mini Cheetah Clone Teardown, By None Other Than Original Designer ”

This is why you should either patent your stuff or make it “as a service”. Bill Gates can buy anything and play any video game he wants. By comparison Richard Stallman can’t even afford decent clothes. If Linux was paid software it would be better.

If Linux was paid software, it wouldn’t be what it is. It’d be Windows, with all the lock-in and “keep up with the rat race” that it includes.

I’ve been using Linux since about 1996. I prefer it to Windows, though I have to use Windows at work.

I prefer the ideals that Stallman and Torvalds represent over the lack of ideals that Gates represents. “Gimme more money” just doesn’t inspire me.

Some people do things to learn. Some people do things to help other people. Some people do things just to get more cash.

It takes all kinds of people to get things done in this world, but I know I prefer those whose main goal isn’t merely to fill their own pockets as full as possible.

Great answer!

To what purpose, what service? This project is the result of a master thesis IIRC, it doesn’t really fill any marketable need. It is a really good applied research project that could be the starting point of other marketable developpements. Leaving it open is a way of boosting innovation, and makes a lot of sense when an organization is there to research and not specifically to make money (even if they are always happy to get more funding). The fact that it’s pretty hard to actually find those clone robots available after a couple of years also shows that the market is not really there. Look at boston dynamics, they are still struggling to market their robots.

In the same way, if Linux was paid software, maybe it wouldn’t exist at all today. In what market Linux would have been competitive at the time? Was there room for a new competitor? I’m not sure, but what I am sure of is that Linux plays a bigger role in computing than any other OS competitor in the current times.

The “as a service” business model is actually an interesting lead on how to market free software: provide knowledge about your system for people to understand it and potentialy improve it or create new uses, while making it easier for a person to pay you regularly instead of deploying it themselves. Of course you have to absorb the NRE costs compared to any competition that could just copy your service, but there are ways to counter that (mainly closing parts of the functionnalities sadly).

As for Stallman and Gates, your comment is funny, but they simply followed different goals. Not everything is about getting more money than needed.

(Also, good luck enforcing a patent in China, your best option is probably to keep everything secret.)

I think the point is that innovative companies like you cited in Boston Dynamics are unlikely to be able to continue to innovate if knock off clones of their products keep appearing from China. You can’t survive if you pay for all the R&D and someone else gets to commoditize your product.

The point of the author in working on his master’s degree probably has something to do with wanting to be compensated at some point to pay for his education. Maybe he does not need to be Bill Gates but he probably would like to eat and be able to afford a nice vacation some day. No one can really afford to be absolutely altruistic when performing their craft.

Indeed, but the mini cheetah wasn’t a development from Boston dynamics. I was only referring to this company to illustrate the lack of existing market. Their work is not open AFAIK, and it potentially won’t make them successful if there are no problems to solve economically with their solutions. Funding this kind of project publicly to open the results may be the best option, in order for future commercial developers to find a way to use (part of) it and market it.

I think we shouldn’t be that afraid of the copying of the results of such projects. They are far from commercial products, and a big part of the valorization of such results actually comes from the know-how of people working on them. If you’ve read Ben Katz’s blog entry, you’ve seen that he immediately noticed some dumb vestigial design aspects that he would have immediately changed. The Chinese didn’t because they don’t understand the project as much as the original team. And if they did and actually changed the design, it would be to the benefit of everyone (ideally, if they respect the licensing terms). In that way, a future company hiring Ben Katz or people having worked on the origonal project would have a significant edge over potential copying competition, at least for a couple of years. I get that it’s not reassuring to investors when you don’t have a nice description of your valued and locked down IP to show them, but it shouldn’t be frightening to engineers.

As for the author of the project, I don’t know about the MIT rules, but I would be extremely surprised if the original author has any ownership on the intellectual property generated by his project. Making the results open may actually be the only way for him to keep using the results of his work…

Also I’m not talking about being altruistic. To reassure you, people working in research center are still paid, just not necessarily to generate directly profitable results. You could also argue that he found retribution in the fame he got from the project, the knowledge he acquired, the entry in the resume and the degree he received. I have no doubt a graduated MIT engineer will find ways to be able to eat and take some vacations ;).

The inescapable fact is that if you open source your design it is a lot harder to make a profit on it. You have all the NRE and research costs and your competitors don’t, hence they can sell it for less and still make a profit. If it is open source anyone can make it themselves. In the case of software it completely removes the ability to profit from it unless you offer some form of technical support. Considering robotics is very software based it isn’t a good idea to open source any of it if you want to make a profit.

In order for open source to be profitable you need to provide something the user can’t provide for themselves. For physical parts that is easy, most people don’t have access to or it is to expensive to have the part machined or injection molded. For software when anyone can download and run it you can’t make much profit, a lot of companies that make open source software ask for donations, that should tell you something about it’s potential profitability.

For software there are certain steps you can take to make it profitable, again to do with providing things that the user can’t, like web hosting or technical support.

You can never avoid the copying problem though. The only way around it is keeping at least part of the system closed source but even then someone might find a way to reverse engineer it and the open source community wouldn’t like it much.

Often universities allow students to patent their designs and can even help the students with the patent process, it all depends on the project proposal and the agreement with the university and any agreements with anyone that funded the project. It absolutely is possible and often encouraged to patent any new ideas you have during your time at university and I know of a few students at my university that have turned a project into a full business, I think the university offers its support for a fraction of the company and staff working at the university are open to create their own deals (some have companies themselves that are independent of the university). The goal of universities is to teach and to do research, none of it needs to be open source

Simply put Ren, people that create as a passion instead of a career drive science and technology.

People that create for a career instead of a passion drive the economy.

Look at the 3d printer world for a good example. The hobbiests drive the industry forward

In what world?

Are hobbyists creating polyjet 3D printers? Are they creating functional metal 3D printers?

I don’t think so, all the advanced and new features tend to come on expensive closed systems first, this even applies somewhat to slicers.

Hobbyists may be driving demand, but for the most part they aren’t driving the industry forward. Most new open features tend to come from research labs or master’s thesis anyway, not really from hobbyists.

If linux was paid software, it would have never achieved adoption and would become a proprietary shitpile even if it did survive ten years (more likely it would be nothing more than a stub article in computing history by now). SaaS is dreck and represents a slide into techno-Lysenkoism. Total drain of competence on a long enough timeline

Well BSD is free and just look at how it’s taking the world by storm. Maybe, being free isn’t enough for success.

The reason why BSD isn’t taking the world by storm is because of the success of Linux, and the success of Linux is due to AT&T’s attacks against BSD in its early days.

iOS and macOS are based on BSD… they haven’t maybe ‘taken the [whole] world by storm’ but I’d say they’re doing pretty well?

If RMS cannot afford decent clothing it’s of his own doing.

My favorite parts is how the product manual talks about their “patented technology” dual absolute encoder and their “three years of painstaking development”. Typical. Even if the original author did not patent the device, they should have at least given a shout out for using his idea. If China will not respect other people’s work they should be removed from the WTO.

Agreed. The CCP should have been economically forced into breaking up after the 1989 Tiananmen Square Massacre, similar to what happened to the Eastern Block in soviet Europe.

They like to pretend they’re “one people one China” but there are 50 or more different ethnic groups and anywhere from 3 to 23 (or more depending on who you ask) different regions. It is past time for those regions to be independent nations. Let Xi the Pooh keep Beijing, the rest of the world can fence it off and treat it like a nature preserve, a living example of what NOT to do.

The sensoring of the actuator is actually partly innovative. But it would have been nice to open it (and maybe mandatory, depending on the license).

About China, I have mixed feelings in this regard. On one hand they can simply rip people’s work and make more margin because of their extremely low cost of production, killing the devs, but on the other hand I had interactions with Chinese companies about having access to their source codes that I wouldn’t dream to have with EU or US companies. In one case, it was because they weren’t respecting an open license, and they fixed the situation immediately. The lack of openness was more due to a lack of care but they were happy to comply. I have yet to encounter such interactions outside of China…

I’m not going to totally defend China’s approach to intellectual property, but what often gets elided in these discussions is that IP isn’t a free-for-all in China, they just have a different system. Bunny Huang has a nice explanation of this: https://www.youtube.com/watch?v=y5QkM2Work0

Remove China from the WTO to protect western laborers and inventors? LMAO what do you think global trade is for? Helping you?

Kinda reminds me of the tendons in a horse’s leg. There’s really not much down there except.. well, biological equivalents of drive belts, tensioners, and pulleys. Which is also why the legs never heal right. They barely even have blood below a certain point.

It is always interesting to me how much history is lost. How can you “clone” something that is open source, and openly shared? Most quadruped research was funded by DARPA, and the research was all done out in the open, and documented in many papers, and git repositories. “Chinese clone” always has a negative connotation?

I’ve written about some of the past here: https://github.com/MAVProxyUser/ConsumerQuadruped

Mind you… these companies that fail to directly credit MIT (until forced) like Unitree, are being unscrupulous to say the least. https://github.com/unitreerobotics/unitree_legged_sdk/issues/55

Leave a Reply Cancel reply

Please be kind and respectful to help make the comments section excellent. ( Comment Policy )

This site uses Akismet to reduce spam. Learn how your comment data is processed .

Never miss a hack

If you missed it.

Mining And Refining: Uranium And Plutonium

Programming Ada: First Steps On The Desktop

The Hunt For MH370 Goes On With Barnacles As A Lead

MXM: Powerful, Misused, Hackable

VCF East 2024 Was Bigger And Better Than Ever

Our columns.

This Week In Security: Cisco, Mitel, And AI False Flags

Keebin’ With Kristina: The One With The Transmitting Typewriter

Supercon 2023: Alex Lynd Explores MCUs In Infosec

FLOSS Weekly Episode 780: Zoneminder — Better Call Randal

Slicing And Dicing The Bits: CPU Design The Old Fashioned Way

By using our website and services, you expressly agree to the placement of our performance, functionality and advertising cookies. Learn more

Online Optimal Landing Control of the MIT Mini Cheetah

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

• Dean’s Office
• External Advisory Council
• Computing Council
• Extended Computing Council
• Undergraduate Advisory Group
• Break Through Tech AI
• Building 45 Event Space
• Infinite Mile Awards: Past Winners
• Frequently Asked Questions
• Undergraduate Programs
• Graduate Programs
• Educating Computing Bilinguals
• Online Learning
• Industry Programs
• AI Policy Briefs
• Envisioning the Future of Computing Prize
• SERC Symposium 2023
• SERC Case Studies
• SERC Scholars Program
• SERC Postdocs
• Common Ground Subjects
• For First-Year Students and Advisors
• For Instructors: About Common Ground Subjects
• Common Ground Award for Excellence in Teaching
• New and Incoming Faculty
• Faculty Resources
• Faculty Openings
• Search for: Search
• MIT Homepage

3 Questions: How the MIT mini cheetah learns to run

CSAIL scientists came up with a learning pipeline for the four-legged robot that learns to run entirely by trial and error in simulation.

It’s been roughly 23 years since one of the first robotic animals trotted on the scene, defying classical notions of our cuddly four-legged friends. Since then, a barrage of the walking, dancing, and door-opening machines have commanded their presence, a sleek mixture of batteries, sensors, metal, and motors. Missing from the list of cardio activities was one both loved and loathed by humans (depending on whom you ask), and which proved slightly trickier for the bots: learning to run. 

Researchers from MIT’s Improbable AI Lab, part of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and directed by MIT Assistant Professor Pulkit Agrawal, as well as the Institute of AI and Fundamental Interactions (IAIFI) have been working on fast-paced strides for a robotic mini cheetah — and their model-free reinforcement learning system broke the record for the fastest run recorded. Here, MIT PhD student Gabriel Margolis and IAIFI postdoc Ge Yang discuss just how fast the cheetah can run.

Q: We’ve seen videos of robots running before. Why is running harder than walking?   

A: Achieving fast running requires pushing the hardware to its limits, for example by operating near the maximum torque output of motors. In such conditions, the robot dynamics are hard to analytically model. The robot needs to respond quickly to changes in the environment, such as the moment it encounters ice while running on grass. If the robot is walking, it is moving slowly and the presence of snow is not typically an issue. Imagine if you were walking slowly, but carefully: you can traverse almost any terrain. Today’s robots face an analogous problem. The problem is that moving on all terrains as if you were walking on ice is very inefficient, but is common among today’s robots. Humans run fast on grass and slow down on ice — we adapt. Giving robots a similar capability to adapt requires quick identification of terrain changes and quickly adapting to prevent the robot from falling over. In summary, because it’s impractical to build analytical (human-designed) models of all possible terrains in advance, and the robot’s dynamics become more complex at high-velocities, high-speed running is more challenging than walking.

Q: Previous agile running controllers for the MIT Cheetah 3 and mini cheetah, as well as for Boston Dynamics’ robots, are “analytically designed,” relying on human engineers to analyze the physics of locomotion, formulate efficient abstractions, and implement a specialized hierarchy of controllers to make the robot balance and run. You use a “learn-by-experience model” for running instead of programming it. Why?

A: Programming how a robot should act in every possible situation is simply very hard. The process is tedious, because if a robot were to fail on a particular terrain, a human engineer would need to identify the cause of failure and manually adapt the robot controller, and this process can require substantial human time. Learning by trial and error removes the need for a human to specify precisely how the robot should behave in every situation. This would work if: (1) the robot can experience an extremely wide range of terrains; and (2) the robot can automatically improve its behavior with experience.

Thanks to modern simulation tools, our robot can accumulate 100 days’ worth of experience on diverse terrains in just three hours of actual time. We developed an approach by which the robot’s behavior improves from simulated experience, and our approach critically also enables successful deployment of those learned behaviors in the real world. The intuition behind why the robot’s running skills work well in the real world is: Of all the environments it sees in this simulator, some will teach the robot skills that are useful in the real world. When operating in the real world, our controller identifies and executes the relevant skills in real-time.

Q: Can this approach be scaled beyond the mini cheetah? What excites you about its future applications?   

A: At the heart of artificial intelligence research is the trade-off between what the human needs to build in (nature) and what the machine can learn on its own (nurture). The traditional paradigm in robotics is that humans tell the robot both what task to do and how to do it. The problem is that such a framework is not scalable, because it would take immense human engineering effort to manually program a robot with the skills to operate in many diverse environments. A more practical way to build a robot with many diverse skills is to tell the robot what to do and let it figure out the how. Our system is an example of this. In our lab, we’ve begun to apply this paradigm to other robotic systems, including hands that can pick up and manipulate many different objects.

This work is supported by the DARPA Machine Common Sense Program, Naver Labs, MIT Biomimetic Robotics Lab, and the NSF AI Institute of AI and Fundamental Interactions. The research was conducted at the Improbable AI Lab.

Related Stories

• April 24, 2024 | Liftoff! NASA’s Next-Generation Solar Sail Boom Technology Launched
• April 24, 2024 | 3.7 Billion Years Old: Oldest Undisputed Evidence of Earth’s Magnetic Field Uncovered in Greenland
• April 24, 2024 | Revolutionizing Renewable Energy: Innovative Salt Battery Efficiently Harvests Osmotic Power
• April 24, 2024 | 30 Times Clearer – Scientists Develop Improved Mid-Infrared Microscope
• April 24, 2024 | Shedding Pounds, Dodging Cancer: The Life-Saving Promise of Bariatric Surgery

How the MIT Mini Cheetah Robot Learns To Run Entirely by Trial and Error

By Rachel Gordon, MIT CSAIL March 29, 2022

MIT’s mini cheetah, using a model-free reinforcement learning system, broke the record for the fastest run recorded. Credit: Photo courtesy of MIT CSAIL.

CSAIL scientists came up with a learning pipeline for the four-legged robot that learns to run entirely by trial and error in simulation.

It’s been roughly 23 years since one of the first robotic animals trotted on the scene, defying classical notions of our cuddly four-legged friends. Since then, a barrage of the walking, dancing, and door-opening machines have commanded their presence, a sleek mixture of batteries, sensors, metal, and motors. Missing from the list of cardio activities was one both loved and loathed by humans (depending on whom you ask), and which proved slightly trickier for the bots: learning to run.

Researchers from MIT ’s Improbable AI Lab, part of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and directed by MIT Assistant Professor Pulkit Agrawal, as well as the Institute of AI and Fundamental Interactions (IAIFI) have been working on fast-paced strides for a robotic mini cheetah — and their model-free reinforcement learning system broke the record for the fastest run recorded. Here, MIT PhD student Gabriel Margolis and IAIFI postdoc Ge Yang discuss just how fast the cheetah can run.

Q: We’ve seen videos of robots running before. Why is running harder than walking?

A: Achieving fast running requires pushing the hardware to its limits, for example by operating near the maximum torque output of motors. In such conditions, the robot dynamics are hard to analytically model. The robot needs to respond quickly to changes in the environment, such as the moment it encounters ice while running on grass. If the robot is walking, it is moving slowly and the presence of snow is not typically an issue. Imagine if you were walking slowly, but carefully: you can traverse almost any terrain. Today’s robots face an analogous problem. The problem is that moving on all terrains as if you were walking on ice is very inefficient, but is common among today’s robots. Humans run fast on grass and slow down on ice — we adapt. Giving robots a similar capability to adapt requires quick identification of terrain changes and quickly adapting to prevent the robot from falling over. In summary, because it’s impractical to build analytical (human-designed) models of all possible terrains in advance, and the robot’s dynamics become more complex at high-velocities, high-speed running is more challenging than walking.

Q: Previous agile running controllers for the MIT Cheetah 3 and mini cheetah, as well as for Boston Dynamics’ robots, are “analytically designed,” relying on human engineers to analyze the physics of locomotion, formulate efficient abstractions, and implement a specialized hierarchy of controllers to make the robot balance and run. You use a “learn-by-experience model” for running instead of programming it. Why?

A: Programming how a robot should act in every possible situation is simply very hard. The process is tedious, because if a robot were to fail on a particular terrain, a human engineer would need to identify the cause of failure and manually adapt the robot controller, and this process can require substantial human time. Learning by trial and error removes the need for a human to specify precisely how the robot should behave in every situation. This would work if: (1) the robot can experience an extremely wide range of terrains; and (2) the robot can automatically improve its behavior with experience.

Thanks to modern simulation tools, our robot can accumulate 100 days’ worth of experience on diverse terrains in just three hours of actual time. We developed an approach by which the robot’s behavior improves from simulated experience, and our approach critically also enables successful deployment of those learned behaviors in the real world. The intuition behind why the robot’s running skills work well in the real world is: Of all the environments it sees in this simulator, some will teach the robot skills that are useful in the real world. When operating in the real world, our controller identifies and executes the relevant skills in real-time.

Q: Can this approach be scaled beyond the mini cheetah? What excites you about its future applications?

A: At the heart of artificial intelligence research is the trade-off between what the human needs to build in (nature) and what the machine can learn on its own (nurture). The traditional paradigm in robotics is that humans tell the robot both what task to do and how to do it. The problem is that such a framework is not scalable, because it would take immense human engineering effort to manually program a robot with the skills to operate in many diverse environments. A more practical way to build a robot with many diverse skills is to tell the robot what to do and let it figure out the how. Our system is an example of this. In our lab, we’ve begun to apply this paradigm to other robotic systems, including hands that can pick up and manipulate many different objects.

This work was supported by the DARPA Machine Common Sense Program, the MIT Biomimetic Robotics Lab, NAVER LABS, and in part by the National Science Foundation AI Institute for Artificial Intelligence Fundamental Interactions, United States Air Force-MIT AI Accelerator, and MIT-IBM Watson AI Lab. The research was conducted by the Improbable AI Lab.

More on SciTechDaily

One Giant Leap for MIT’s Robotic Mini Cheetah

Untethered Electric Robotic Cheetah Runs and Jumps Across the Grass

Shape-Shifting, Free-Roaming Soft Robot Created by Stanford Engineers [Video]

RoboGrammar: Automated & Optimized Robot Design for Traversing Various Terrain Types

What to Expect When You’re Expecting Robots: The Future of Human-Robot Collaboration

How To Help Humans Understand Robots – To Collaborate Faster and More Effectively

Two-Legged Robot Mimics Human Balance While Running and Jumping [Video]

New Machine-Learning System Gives Robots Social Skills

Be the first to comment on "how the mit mini cheetah robot learns to run entirely by trial and error", leave a comment cancel reply.

Email address is optional. If provided, your email will not be published or shared.

Save my name, email, and website in this browser for the next time I comment.

MIT MechE Featured Alumni

In November 2018, Professor Sangbae Kim brought the mini cheetah robot onto The Tonight Show’s “Tonight Show-botics” segment. Much to the delight of host Jimmy Fallon, the mini cheetah did some yoga, got back up after falling, and executed a perfect backflip. Behind the stage, Benjamin Katz '16, SM '18 was remotely controlling the cheetah’s nimble maneuvers.

For Katz, waiting in the wings as the robot performed in front of a national audience was the culmination of nearly five years of work.

As an undergraduate at MIT, Katz studied mechanical engineering, opting for the flexible Course 2A degree program with a concentration in controls, instrumentation, and robotics. Toward the end of his first year, he emailed Kim to see if there were any job opportunities in Kim’s Biomimetic Robotics Lab. He then spent the summer in Kim’s lab as part of the MIT UROP – or Undergraduate Research Opportunities Program. For his UROP research and undergraduate thesis, he began to look at how to utilize pieces built for the electronics hobby market in robotics. “You can find really high-performance motors built for things like remote control airplanes and drones. I basically thought you could also use these parts for robots, which is something no one was doing,” recalls Katz.

Kim was immediately impressed by Katz’s abilities an engineer and designer.

“Ben is an extremely versatile engineer who can cover structure and mechanism design, electric motor dynamics, power electronics, and classical control, a range of expertise usually requiring four-to-five engineers to cover,” says Kim.

After deciding to pursue a master’s degree in mechanical engineering at MIT, Katz continued working in Kim’s lab and developed solutions for actuators in robotics. While working on the third iteration of Kim’s robot, known as Cheetah 3, Katz and his labmates shifted their focus to developing a smaller version of the robot.

MIT-Mini-Cheetah-1024-.jpg

“There are a lot of nice things about having a smaller robot: if something breaks you can easily fix it, it’s cheaper, and it’s safe enough for one person to wrangle alone,” says Katz. “Even though a small robot may not always be the most practical for real-world applications, its controllers, software, and research can be trivially ported to a big robot that can carry larger payloads.”

Drawing upon his undergraduate research, Katz and the research team used twelve motors originally designed for drones to build actuators in each joint of the small quadruped robot that would be dubbed the “mini cheetah.”

Armed with this smaller robot, Katz set out to make the mini cheetah more agile and resilient. Alongside then EECS student Jared Di Carlo ‘19, MNG ‘20, Katz focused on controls related to locomotion in the mini cheetah. In class 6.832, Underactuated Robotics, taught by Professor Russ Tedrake, the pair worked on a project that would allow the mini cheetah safely backflip from a crouched position.

“It was basically a giant offline optimization problem to get the mini cheetah to backflip,” says Katz.

Using offline non-linear optimization to generate the backflip trajectory, he and Di Carlo were able to program the mini cheetah to crouch and rotate 360 degrees around an axis.

While working on the cheetah, Katz was constantly pursuing other engineering projects as a hobby. This included a very different rotating robot as a pet project. Alongside Di Carlo, Katz utilized the MIT community makerspace known as MITERS to develop a robot that could solve a Rubik’s Cube in a record-breaking 0.38 seconds.

“That project was purely for fun during MIT’s Independent Activities Period,” recalls Katz. “We used custom-built actuators on each of the Rubik’s Cube’s faces alongside webcams to identify the colors and move the blocks accordingly.”

He chronicled his other pet projects on his “build-its” blog, which developed a strong following. Projects included planar magnetic headphones, a desktop furuta pendulum, and an electric travel ukulele.

“Ben was constantly building and analyzing something along with our lab and class projects during his entire time at MIT,” says Kim. “His incessant desire to learn, build, and analyze is quite remarkable.”

After graduating with his master’s degree in 2018, Katz worked as a technical associate in Kim’s lab before accepting a position at Boston Dynamics in 2019.

As a designer at Boston Dynamics, Katz has transitioned from cheetah robots to humanoid robots on ATLAS, a research platform billed as the “world’s most dynamic humanoid robot.” Much like the mini cheetah, ATLAS can execute incredibly dynamic maneuvers including backflips and even parkour.

While the mini cheetah holding yoga poses and ATLAS doing parkour seems like entertainment befitting The Tonight Show, Katz is quick to remind others that these robots are fulfilling a real-world need. The robots could someday maneuver in areas that are too dangerous for humans – including buildings that are on fire and disaster areas. They could open new possibilities for life-saving disaster relief and first-responders in emergencies.

“What we did in Sangbae’s lab is going to help make these machines ubiquitous and actually useful in the real world as viable products,” adds Katz.

Other Alumni Profiles

Jump to navigation

3 Questions: How the MIT mini cheetah learns to run

Photo courtesy of MIT CSAIL.

MIT’s mini cheetah, using a model-free reinforcement learning system, broke the record for the fastest run recorded. The team is led by LIDS Affiliate Member Pulkit Agrawal.

Article Author

Rachel Gordon | MIT CSAIL

Date Published

March 17, 2022

Link to Original Article

It’s been roughly 23 years since one of the first robotic animals trotted on the scene, defying classical notions of our cuddly four-legged friends. Since then, a barrage of the walking, dancing, and door-opening machines have commanded their presence, a sleek mixture of batteries, sensors, metal, and motors. Missing from the list of cardio activities was one both loved and loathed by humans (depending on whom you ask), and which proved slightly trickier for the bots: learning to run. 

Researchers from MIT’s Improbable AI Lab, part of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and directed by MIT Assistant Professor Pulkit Agrawal, as well as the Institute of AI and Fundamental Interactions (IAIFI) have been working on fast-paced strides for a robotic mini cheetah — and their model-free reinforcement learning system broke the record for the fastest run recorded. Here, MIT PhD student Gabriel Margolis and IAIFI postdoc Ge Yang discuss just how fast the cheetah can run.  

Q: We’ve seen videos of robots running before. Why is running harder than walking?   

A: Achieving fast running requires pushing the hardware to its limits, for example by operating near the maximum torque output of motors. In such conditions, the robot dynamics are hard to analytically model. The robot needs to respond quickly to changes in the environment, such as the moment it encounters ice while running on grass. If the robot is walking, it is moving slowly and the presence of snow is not typically an issue. Imagine if you were walking slowly, but carefully: you can traverse almost any terrain. Today’s robots face an analogous problem. The problem is that moving on all terrains as if you were walking on ice is very inefficient, but is common among today’s robots. Humans run fast on grass and slow down on ice — we adapt. Giving robots a similar capability to adapt requires quick identification of terrain changes and quickly adapting to prevent the robot from falling over. In summary, because it’s impractical to build analytical (human-designed) models of all possible terrains in advance, and the robot's dynamics become more complex at high-velocities, high-speed running is more challenging than walking.

Q: Previous agile running controllers for the MIT Cheetah 3 and mini cheetah, as well as for Boston Dynamics’ robots, are “analytically designed,” relying on human engineers to analyze the physics of locomotion, formulate efficient abstractions, and implement a specialized hierarchy of controllers to make the robot balance and run. You use a “learn-by-experience model” for running instead of programming it. Why? 

A: Programming how a robot should act in every possible situation is simply very hard. The process is tedious, because if a robot were to fail on a particular terrain, a human engineer would need to identify the cause of failure and manually adapt the robot controller, and this process can require substantial human time. Learning by trial and error removes the need for a human to specify precisely how the robot should behave in every situation. This would work if: (1) the robot can experience an extremely wide range of terrains; and (2) the robot can automatically improve its behavior with experience. 

Thanks to modern simulation tools, our robot can accumulate 100 days’ worth of experience on diverse terrains in just three hours of actual time. We developed an approach by which the robot’s behavior improves from simulated experience, and our approach critically also enables successful deployment of those learned behaviors in the real world. The intuition behind why the robot’s running skills work well in the real world is: Of all the environments it sees in this simulator, some will teach the robot skills that are useful in the real world. When operating in the real world, our controller identifies and executes the relevant skills in real-time.  

Q: Can this approach be scaled beyond the mini cheetah? What excites you about its future applications?   

A: At the heart of artificial intelligence research is the trade-off between what the human needs to build in (nature) and what the machine can learn on its own (nurture). The traditional paradigm in robotics is that humans tell the robot both what task to do and how to do it. The problem is that such a framework is not scalable, because it would take immense human engineering effort to manually program a robot with the skills to operate in many diverse environments. A more practical way to build a robot with many diverse skills is to tell the robot what to do and let it figure out the how. Our system is an example of this. In our lab, we’ve begun to apply this paradigm to other robotic systems, including hands that can pick up and manipulate many different objects.

This work is supported by the DARPA Machine Common Sense Program, Naver Labs, MIT Biomimetic Robotics Lab, and the NSF AI Institute of AI and Fundamental Interactions. The research was conducted at the Improbable AI Lab.

Reprinted with permission of MIT News.

• Companies Founded by LIDS Community Members
• LIDS/All Magazine
• Directions to LIDS Offices
• LIDS Overview
• Statistical Inference and Machine Learning
• Optimization and Game Theory
• Systems Theory, Control, and Autonomy
• Aerospace Controls Laboratory (ACL)
• Communications and Networking Research Group (CNRG)
• Data to AI Group (DAI)
• Electric Energy Systems Group (EESG@MIT)
• Inference and Stochastic Networks Group (ISNG)
• Stochastic Systems Group (SSG)
• Wireless Information and Network Sciences Laboratory (WINSLab)
• Event Calendar
• LIDS Seminar Series
• Full Archive
• Spring 2014
• Spring 2015
• Spring 2016
• Spring 2017
• Spring 2018
• Spring 2019
• Spring 2020
• Spring 2021
• Spring 2022
• LIDS & Stats Tea
• Spring 2023
• Conferences and Workshops
• Faculty and PIs
• Administrative Staff
• Research Staff
• Information for Visitors and New Members
• LIDS Research Group Snapshots
• Book a Conference Room
• Printer Instructions for Windows
• Printing Instructions for Mac
• Printing Instructions for Linux
• LIDS Theses & Publications
• Request Repairs
• Ship & Receive Mail
• Travel Rules and Reimbursements
• Subscribing to Calendar Feeds
• Using Vibe boards

Contact Us: 617-253-2142

Laboratory for Information and Systems Decisions

Massachusetts Institute of Technology 77 Massachusetts Avenue Room 32-D608 Cambridge, MA 02136

Suggestions or feedback?

MIT News | Massachusetts Institute of Technology

• Machine learning
• Social justice
• Black holes
• Classes and programs

Departments

• Aeronautics and Astronautics
• Brain and Cognitive Sciences
• Architecture
• Political Science
• Mechanical Engineering

Centers, Labs, & Programs

• Abdul Latif Jameel Poverty Action Lab (J-PAL)
• Picower Institute for Learning and Memory
• Lincoln Laboratory
• School of Architecture + Planning
• School of Engineering
• School of Humanities, Arts, and Social Sciences
• Sloan School of Management
• School of Science
• MIT Schwarzman College of Computing

From “cheetah-noids” to humanoids

Press contact :.

Previous image Next image

In November 2018, MIT Professor Sangbae Kim brought his mini cheetah robot onto “The Tonight Show’s” Tonight Show-botics segment. Much to the delight of host Jimmy Fallon, the mini cheetah did some yoga, got back up after falling, and executed a perfect backflip. Behind the stage, Benjamin Katz '16, SM '18 was remotely controlling the cheetah’s nimble maneuvers.

For Katz, waiting in the wings as the robot performed in front of a national audience was the culmination of nearly five years of work.

As an undergraduate at MIT, Katz studied mechanical engineering, opting for the flexible Course 2A degree program with a concentration in controls, instrumentation, and robotics. Toward the end of his first year, he emailed Kim to see if there were any job opportunities in Kim’s Biomimetic Robotics Lab. He then spent the summer in Kim’s lab as part of the MIT Undergraduate Research Opportunities Program (UROP). For his UROP research and undergraduate thesis, he began to look at how to utilize pieces built for the electronics hobby market in robotics. “You can find really high-performance motors built for things like remote control airplanes and drones. I basically thought you could also use these parts for robots, which is something no one was doing,” recalls Katz.

Kim was immediately impressed by Katz’s abilities an engineer and designer.

“Ben is an extremely versatile engineer who can cover structure and mechanism design, electric motor dynamics, power electronics, and classical control, a range of expertise usually requiring four-to-five engineers to cover,” says Kim.

After deciding to pursue a master’s degree in mechanical engineering at MIT, Katz continued working in Kim’s lab and developed solutions for actuators in robotics. While working on the third iteration of Kim’s robot, known as Cheetah 3, Katz and his labmates shifted their focus to developing a smaller version of the robot.

“There are a lot of nice things about having a smaller robot: If something breaks you can easily fix it, it’s cheaper, and it’s safe enough for one person to wrangle alone,” says Katz. “Even though a small robot may not always be the most practical for real-world applications, its controllers, software, and research can be trivially ported to a big robot that can carry larger payloads.”

Drawing upon his undergraduate research, Katz and the research team used 12 motors originally designed for drones to build actuators in each joint of the small quadruped robot that would be dubbed the “mini cheetah.”

Armed with this smaller robot, Katz set out to make the mini cheetah more agile and resilient. Alongside then-EECS student Jared Di Carlo '19, Mng '20, Katz focused on controls related to locomotion in the mini cheetah. In class 6.832 (Underactuated Robotics), taught by Professor Russ Tedrake, the pair worked on a project that would allow the mini cheetah safely backflip from a crouched position.

“It was basically a giant offline optimization problem to get the mini cheetah to backflip,” says Katz.

Using offline nonlinear optimization to generate the backflip trajectory, he and Di Carlo were able to program the mini cheetah to crouch and rotate 360 degrees around an axis.

While working on the cheetah, Katz was constantly pursuing other engineering projects as a hobby. This included a very different rotating robot as a pet project. Alongside Di Carlo, Katz utilized the MIT community makerspace known as MITERS to develop a robot that could solve a Rubik’s Cube in a record-breaking 0.38 seconds.

“That project was purely for fun during MIT’s Independent Activities Period,” recalls Katz. “We used custom-built actuators on each of the Rubik’s Cube’s faces alongside webcams to identify the colors and move the blocks accordingly.”

He chronicled his other pet projects on his “build-its” blog, which developed a strong following. Projects included planar magnetic headphones, a desktop Furuta pendulum, and an electric travel ukulele.

“Ben was constantly building and analyzing something along with our lab and class projects during his entire time at MIT,” says Kim. “His incessant desire to learn, build, and analyze is quite remarkable.”

After graduating with his master’s degree in 2018, Katz worked as a technical associate in Kim’s lab before accepting a position at Boston Dynamics in 2019.

As a designer at Boston Dynamics, Katz has transitioned from cheetah robots to humanoid robots on ATLAS, a research platform billed as the “world’s most dynamic humanoid robot.” Much like the mini cheetah, ATLAS can execute incredibly dynamic maneuvers, including backflips and even parkour.

While the mini cheetah holding yoga poses and ATLAS doing parkour seems like entertainment befitting "The Tonight Show," Katz is quick to remind others that these robots are fulfilling a real-world need. The robots could someday maneuver in areas that are too dangerous for humans — including buildings that are on fire and disaster areas. They could open new possibilities for lifesaving disaster relief and first-responders in emergencies.

“What we did in Sangbae’s lab is going to help make these machines ubiquitous and actually useful in the real world as viable products,” adds Katz.

Share this news article on:

Related links.

• Sangbae Kim
• Biomimetic Robotics Laboratory
• Department of Mechanical Engineering

Related Topics

• Mechanical engineering
• Artificial intelligence
• Maker movement
• Undergraduate Research Opportunities Program (UROP)

Related Articles

One giant leap for the mini cheetah

Mini cheetah is the first four-legged robot to do a backflip

Cheetah III robot preps for a role as a first responder

Previous item Next item

More MIT News

Two from MIT awarded 2024 Paul and Daisy Soros Fellowships for New Americans

Read full story →

MIT Emerging Talent opens pathways for underserved global learners

The MIT Edgerton Center’s third annual showcase dazzles onlookers

3 Questions: A shared vocabulary for how infectious diseases spread

Seven from MIT elected to American Academy of Arts and Sciences for 2024

Two MIT teams selected for NSF sustainable materials grants

• More news on MIT News homepage →

Massachusetts Institute of Technology 77 Massachusetts Avenue, Cambridge, MA, USA

• Map (opens in new window)
• Events (opens in new window)
• People (opens in new window)
• Careers (opens in new window)
• Accessibility
• Social Media Hub
• MIT on Facebook
• MIT on YouTube
• MIT on Instagram

MIT Libraries home DSpace@MIT

• DSpace@MIT Home
• MIT Open Access Articles

MIT Cheetah 3: Design and Control of a Robust, Dynamic Quadruped Robot

Open Access Policy

Creative Commons Attribution-Noncommercial-Share Alike

Terms of use

Date issued, collections.

Navigation Menu

Search code, repositories, users, issues, pull requests..., provide feedback.

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly.

To see all available qualifiers, see our documentation .

• Notifications

mit-biomimetics/Cheetah-Software

Folders and files, repository files navigation, cheetah-software.

This repository contains the Robot and Simulation software project. For a getting started guide, see the documentation folder.

The common folder contains the common library with dynamics and utilities The resources folder will contain data files, like CAD of the robot used for the visualization The robot folder will contain the robot program The sim folder will contain the simulation program. It is the only program which depends on QT. The third-party will contain small third party libraries that we have modified. This should just be libsoem for Cheetah 3, which Pat modified at one point.

To build all code:

If you are building code on your computer that you would like to copy over to the mini cheetah, you must replace the cmake command with

otherwise it will not work. If you are building mini cheetah code one the mini cheetah computer, you do not need to do this.

This build process builds the common library, robot code, and simulator. If you just change robot code, you can simply run make -j4 again. If you change LCM types, you'll need to run cmake ..; make -j4 . This automatically runs make_types.sh .

To test the common library, run common/test-common . To run the robot code, run robot/robot . To run the simulator, run sim/sim .

Part of this build process will automatically download the gtest software testing framework and sets it up. After it is done building, it will produce a libbiomimetics.a static library and an executable test-common . Run the tests with common/test-common . This output should hopefully end with

Run simulator

To run the simulator:

• Open the control board
• In the another command window, run the robot control code

3: Cheetah 3, m: Mini Cheetah s: simulation, r: robot

Run Mini cheetah

• Create build folder mkdir mc-build
• Build as mini cheetah executable cd mc-build; cmake -DMINI_CHEETAH_BUILD=TRUE ..; make -j
• Connect to mini cheetah over ethernet, verify you can ssh in
• Copy program to mini cheetah with ../scripts/send_to_mini_cheetah.sh
• ssh into the mini cheetah ssh [email protected]
• Enter the robot program folder cd robot-software-....
• Run robot code ./run_mc.sh

Dependencies:

• Qt 5.10 - https://www.qt.io/download-qt-installer
• LCM - https://lcm-proj.github.io/ (Please make it sure that you have a java to let lcm compile java-extension together)
• Eigen - http://eigen.tuxfamily.org
• mesa-common-dev
• freeglut3-dev
• libblas-dev liblapack-dev

To use Ipopt, use CMake Ipopt option. Ex) cmake -DIPOPT_OPTION=ON ..

Contributors 9

• Python 2.2%
• MATLAB 1.5%
• Makefile 0.1%