Georgia Tech teaches humanoid robots to walk on rough ground ^{CEFR B2}

Researchers at Georgia Tech developed a new training framework that speeds up and lowers the cost of teaching a humanoid robot to walk on varied real-world terrain. The team, led by machine learning PhD student Feiyang Wu, tested a whole-body controller on sand, soggy grass, gravel, slopes, stairs and level ground found on campus or easily simulated.

The work rethinks a common reinforcement learning setup known as teacher and student learning. In the conventional approach a simulated teacher is trained first using detailed simulation information; it explores and learns, then teaches a student that must operate with the limited sensors of a real robot. The researchers identified two problems: sequential training is slow and it can discard useful information gathered by the teacher, and running many simulations requires long GPU compute time.

The Georgia Tech team trained the teacher and student simultaneously so the teacher gradually instructs the student while continuing to learn. The teacher also absorbs data produced by the student, which helps close the teacher–student imitation gap where the student lacks information the teacher used. After simulation training the controller was deployed on a physical humanoid in Ye Zhao's lab. The robot walked smoothly across varied surfaces and adapted when researchers pushed and pulled it. The "Learn to Teach" framework is designed to be generic and can apply to other robot designs and tasks. The work was supported by the Office of Naval Research, the US Department of Agriculture, and the National Science Foundation, and was presented at the IEEE International Conference on Robotics and Automation.