A team of researchers from Facebook AI, UC Berkeley and Carnegie Mellon University’s School of Computer Science announced a new AI technique called Rapid Motor Adaptation (RMA) on Friday. Using this breakthrough in AI, legged robots can now adapt to their environments in real-time.
Until now, legged robots were either fully hand-ceded for the environments they were supposed to function in or were taught to navigate environments through a combination of hand-coding and learning techniques.
RMA, however, is being hailed as the first entirely learning-based system that allows a legged robot to adapt to its environment without previous training or input.
Simulation training and instantaneous adaptation
In the real world, conditions are never static, and hence, robots trained for one situation can’t really perform well outside of those parameters. RMA handles this using two different systems — a base policy and an adaptation module.
The Base policy is where the robot learns how to adapt to different situations using information from different environments, including friction and payload weight. The team sets different variables for the environment, such as a higher incline or lower friction, to let the robot learn the right controls for the right situations.
And since the research team knows the actual conditions the robot encounters in simulation, they can use that information and learning to train the adaptation module to predict them and behave accordingly. The base policy runs at a higher speed as compared to the adaptation module allowing the robot to move around freely without any fine-tuning.
Using this combination of the two subsystems, the robot can adapt to new situations in a matter of seconds. Compared to earlier RL-trained robots, which required several minutes and sometimes even human intervention, this is a huge step up in the field.
During the trails, the robot was able to walk on sand, in mud, on hiking trails, in tall grass, and over a dirt pile without failure. The paper also notes that the robot was able to maintain its height with a 12kg payload with a “high success rate”, successfully walked down steps along a hiking trail in 70% of the trials and was able to navigate a cement pile as well as a pebble pile 80% of the trails.