Fast running is possible only if the hardware can be pushed to its limits. For example, you need to operate at the maximum torque output of motors. These conditions make it difficult to model the dynamics of robots. The robot must be able to react quickly to changes in its environment. For example, if it runs on grass and encounters ice, it will need to move quickly. The robot will walk slowly if it is moving, and snow is usually not an issue. Imagine walking slowly but carefully. You can cross almost any terrain. Robots today face an analogous problem. It is inefficient to move on all terrains like you would on ice, which is a problem that robots today have to deal with. Humans adapt by running fast on grass and slower on ice. Robots need to be able to quickly identify terrain changes and adapt to avoid falling. High-speed running is more difficult than walking because it is impossible to create analytic (human-designed models) of all terrains.
The previous agile running controllers for the MIT Cheetah 3 & mini cheetah as well as Boston Dynamics’ robots were “analytically designed” and relied on human engineers to analyze the physics and formulate efficient abstractions. They then implemented a specialized hierarchy to control the robot’s balance and run. Instead of programming the robot, you use a “learn by experience model” to run it. Why?
It is difficult to program a robot in all possible situations. This is a tedious process because, in the event that a robot fails on a specific terrain, a human engineer would have to identify the problem and modify the robot controller manually. This can take a lot of human time. It is possible to learn by trial and error, which eliminates the need to have a human specify how the robot should behave in each situation. This is possible if the robot can be trained to adapt to different terrains and can also learn from its mistakes.
Modern simulation tools allow our robot to accumulate over 100 days of experience on different terrains in three hours. Our approach improves the robot’s behavior from simulation experience. It also allows for successful deployment of learned behaviors in real life. It is clear that the robot’s skills in running in real-world situations are based on the fact that it can learn from the simulator only those skills that will be useful in real life. Our controller is able to identify and execute the necessary skills when operating in the real-world. This is artificial intelligence research at its core. Robotics has been based on the idea that humans can tell robots what tasks to perform and how to accomplish them. This framework is not scaleable because it would require a lot of human engineering effort to program a robot that can operate in many different environments. It is possible to create a robot with multiple skills by telling it what to do, and then letting it figure out how to do it. This is how our system works. We have begun to apply this principle to robotic systems in our laboratory, such as hands that can manipulate many objects.
