Wu also tailored the training so that the teacher robot learned from what the student robots experienced in the simulation. This helps close what robotics researchers call the teacher-student imitation gap. Essentially, even with proper guidance from the teacher, students are still making decisions based on partial information. When students encounter obstacles or terrain, they may not have the critical data to properly navigate them.
“This gap can be very troubling and difficult to resolve. We sought to alleviate it by having teachers also learn from data collected by students,” Wu said. “Teachers will experience what is possible for their students and we hope that will help them provide better instruction.”
After applying the new approach and training the controller in simulation, they deployed it to a bipedal humanoid robot in Ye Zhao’s lab. It worked well and the robot was able to walk smoothly on various surfaces.
The researchers also tried forcing the robot to push and pull to see if it interfered with its walking, but the robot adapted and adjusted.
Although Wu and his colleagues used bipedal humanoid robots in their experiments, his “Learn to Teach” training framework is designed to be general. Can also be used for other robots with other configurations. It can also be applied to other types of tasks besides walking.
Zhao, CSE assistant professor Anqi Wu and Wu’s co-supervisor, said the control system performed better than the controller provided by the robot’s manufacturer.
He said Wu brings a unique perspective to the robotics work because of his background in machine learning.
“Most of the PhD students in my lab have a background in robotics and control. Feiyan is starting to explore robotics problems from a more theoretical, algorithm-focused background, but it’s not easy,” said Zhao, ME associate professor and Woodruff faculty fellow. “There’s a big barrier. Once students get used to programming or writing math, they might not be motivated to explore working with real hardware. Hiyo has a strong motivation to explore things on both fronts, which is very unique.”
