The past few decades have seen robots go from lumbering behemoths that could barely take a step to nimble contraptions that can do backflips and dance. A lot of work goes into teaching robots how to move, so researchers are increasingly turning to AI to do the heavy lifting, and a team at MIT has seen particularly speedy results. The “Mini Cheetah” designed at the Computer Science and Artificial Intelligence Laboratory (CSAIL) has set a new personal best for movement speed with the help of learning algorithms that ensure it learns from every fall.
Most of the robots you’ve seen placidly walking around on level surfaces have been programmed to do that. The same goes for Boston Dynamics with all its fancy robot demo videos — programmers have written code that tells the robots how to move in order to walk, jump, or do a backflip. However, you can’t account for every possible situation or surface even with an army of programmers. A neural network can be used to leverage a robot’s experience, something we biological organisms do all the time.
After years of walking on uneven surfaces, your brain knows what to do if you lose traction. You might not always save yourself from a tumble, but you can learn and adjust your behavior next time. The MIT team managed something similar for the Mini Cheetah using AI and modern simulation tools. The robot can gain 100 days worth of experience in just three hours of actual time. Through trial and error, the robot gets better at navigating the world, and no one has to explicitly tell it how to do that.
The cheetah’s neural network allowed it to reach its highest speed yet: 3.9 meters per second, or about 8.7 miles per hour. That’s a touch faster than the average human can run. More interestingly, the robot can adjust its gait as it encounters different surfaces and obstacles. The robot looks less elegant in this mode, scrambling around in a way that looks frantic and uncoordinated, but turns out to be more efficient. The video above shows several examples of Mini Cheetah successfully coping with changing conditions while a graceful human-programmed version falls flat.
The team sees this approach to robotics as a necessity for the future. It’s a trade-off between nature and nurture, according to MIT PhD student Gabriel Margolis and postdoc Ge Yang. They note that the same system could teach robots to handle many different tasks. We just have to get used to not telling robots both what to do and how to do it — that framework, they say, is not scalable. In the future we’ll need to let robots learn, and Mini Cheetah is out in front.