SUGARv2 trains humanoid robots in varied whole-body skills from synthetic references

Researchers led by Tianshu Wu and Hao Dong have described SUGARv2 in an arXiv paper, a framework for teaching humanoid robots skills that combine walking and handling objects, guided by vision. To cut the human effort of collecting demonstrations, it adapts reusable task templates to different layouts and synthesises coordinated motions of robot and object as references.

The robot learns from these references with the same imitation reward across tasks: privileged teacher policies are trained first, then visual policies through distillation and reinforcement learning, with varied objects and randomised visuals. Across five tasks it reached a mean success rate of 91.0% in simulation, and the authors report successful transfer to a real humanoid robot.