Quadrupedal robots are becoming more common, but engineers are still figuring out what these machines are capable of. According to a group of MIT researchers, one method to improve their functionality could be to employ AI to help teach the bots how to walk and run.
When engineers design the software that governs the movement of legged robots, they usually write a set of rules for how the machine should respond to various inputs. As an example, if a robot’s sensors detect x amount of force on leg y, it will respond by activating motor a to apply torque b, and so on. Although coding these characteristics is difficult and time-consuming, it provides researchers with precise and predictable control over the robots.
Another option is to employ machine learning, specifically a method known as reinforcement learning, which works through trial and error. Giving your AI model a goal known as a « reward function » (e.g., go as fast as you can) and then releasing it loose to figure out how to attain that end from scratch is how this works. This takes a long time, but it helps if you let the AI to experiment in a simulated environment where time may be sped up. This is why reinforcement learning, or RL, is a popular method for creating AI that can play video games.
This is the method used by MIT engineers to develop new software (known as a « controller ») for the university’s research quadruped, Mini Cheetah. They were able to achieve a new top speed for the robot of 3.9m/s, or around 8.7mph, using reinforcement learning. You can see an example of this in the video below:
Mini Cheetah’s new running gait is, as you can see, a tad awkward. In truth, it appears to be a puppy scrabbling to get moving on a wooden floor. However, according to MIT PhD student Gabriel Margolis (a co-author of the study alongside postdoc fellow Ge Yang), this is because the AI is just optimizing for speed.
« RL finds one method to run swiftly, but given an unspecified reward function, it has no incentive to choose a ‘natural-looking’ or favoured by humans gait, » Margolis writes in an email to The Verge. He claims that while the model may be programmed to generate a more fluid type of mobility, the goal of the project is to optimize for speed alone.
According to Margolis and Yang, one major advantage of employing AI to design controller software is that it takes less time than fiddling about with all the physics. « It’s simply incredibly difficult to program how a robot should behave in every potential situation. » « The method is time-consuming because if a robot fails on a specific terrain, a human engineer must discover the cause of failure and manually change the robot controller, » they explain.
Engineers may use a simulator to set the robot in any number of virtual conditions, from solid pavement to slick rubble, and let it figure things out for itself. Indeed, the MIT team claims that its simulator was able to complete 100 days of staggering, walking, and running in just three hours of real time.
Some businesses that make legged robots are already employing these techniques to create new controllers. Others, such as Boston Dynamics, appear to use more traditional approaches. (This is understandable given the company’s concentration in producing very particular motions, such as the jumps, vaults, and flips featured in its choreographed videos.)
There are other faster-moving robots on the market. Boston Dynamics’ Cheetah bot now holds the quadruped speed record, clocking 28.3 mph – quicker than Usain Bolt. Cheetah, on the other hand, is not only a considerably larger and more powerful machine than MIT’s Mini Cheetah, but it also achieved its record while running on a treadmill and attached to a lever for stability. Without these advantages, AI could be able to compete with the machine.