MIT’s robot cheetah taught itself how to run fast and traverse tricky terrain
What just happened? Scientists at MIT’S Computer Science and Artificial Intelligence Laboratory (CSAIL) have trained a robotic cheetah to break the record for the fastest run ever recorded. The secret was to let the robot figure out how to run through trial and error rather than relying on human engineers to program the bot.
As MIT PhD student Gabriel Margolis and IAIFI postdoc Ge Yang explained in a recent interview, the traditional paradigm in robotics is for humans to tell a robot what to do and how to do it. The problem with that approach is that it isn’t scalable due to the sheer amount of human hours needed to manually program a robot to operate in many different 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.”
One way to get around that limitation is with simulation and AI / machine learning. Using modern simulation tools, the team’s robot was able to accumulate 100 days’ worth of experience on diverse terrains like ice and gravel in just three hours of real time.
The learn-by-experience, or reinforcement, model is on full display in MIT’s latest video and the results are incredibly impressive.
The bot hit a top speed of 3.9 meters per second, or roughly 8.7 mph, when sprinting. Even more impressive is its handling of dicey terrain like gravel. With the human-designed controller, the bot struggles to traverse gravel and even trips and falls when trying to move to the sidewalk. The unit with the learned controller handles the situation effortlessly.