Facebook AI Researchers Create A Robot Able To Adapt To Rapidly Changing Conditions

Work is being performed all around the world to give robots more human-like abilities. One of the key human-like abilities that researchers want robots to have is the ability to react in real-time to environmental changes. Humans, for instance, can remain stable and upright when the surface they are walking on suddenly and unexpectedly changes.

Significant changes in walking service often cause robots to fall over. Facebook AI researchers are working to give robots the ability to adapt to whatever surface they encounter as they move, even if the robot has never experienced those conditions before. The robot must change in real-time to suit the surface it is walking on to prevent the robot from falling and potentially damaging itself.

Researchers from Facebook AI, UC Berkeley, and Carnegie Mellon University School of Computer Science announced Rapid Motor Adaptation (RMA). RMA is a breakthrough in AI enabling legged robots to adapt intelligently and in real-time to unfamiliar terrain and circumstances. RMA harnesses a combination of two policies that were both learned entirely in simulation using a base policy trained through reinforcement learning (RL) and an adaptation module trained using supervised learning.

With RMA, the robot gains an aptitude fundamental to intelligent agents with the ability to adapt to factors in its environment, such as the weight of a backpack unexpectedly placed on the robot or the amount of friction between it and a new surface. The key to this capability is that the robot can adapt without depending on visual input. Until the RMA breakthrough, robots had to be hand-coded to function in the environment they would inhabit or taught to navigate the environment through a combination of hand-coding and learning techniques.

Researchers say that RMA is the first entirely learning-based system that enables legged robots to adapt to their environment by exploring and interacting with the world. Testing showed that robots utilizing RMA outperformed alternative systems when walking on different surfaces, slopes, and obstacles. They also performed better when given different payloads to carry.