Georgia Tech teaches humanoid robots to walk on rough ground ^{CEFR B1}

Researchers at Georgia Tech have developed a faster, less expensive way to train a humanoid robot to walk on uneven, real-world terrain. PhD student Feiyang Wu led the work on a new whole-body controller, and the team tested it on sand, soggy grass, gravel, slopes, stairs and level ground found on campus or simulated.

The project rethinks a reinforcement learning approach called teacher and student learning. In the standard method a simulated teacher is trained first with detailed simulation data and then teaches a student with more limited information. The team found this sequence takes a long time and can waste useful data, because simulations require many hours on GPU chips.

Their solution trains teacher and student at the same time. The teacher both teaches the student and learns from data produced by the student, which reduces the imitation gap. After simulation training the controller ran on a physical humanoid in Ye Zhao's lab, where the robot walked smoothly and adapted when researchers pushed and pulled it. Wu presented the framework at the IEEE International Conference on Robotics and Automation, and the work had support from US research agencies.