Lizhi Yang
Lizhi Yang is a Mechanical Engineering PhD candidate at the California Institute of Technology (Caltech) in Pasadena, where he has been conducting robotics research since 2022. He is best known for his work on safety-critical reinforcement learning, particularly the PAC-MAN framework that combines control barrier functions with RL to enable whole-body safety for humanoid robots in adversarial scenarios such as dodgeball. His research focuses on the intersection of control theory, machine learning, and robotic safety, including perception-aware methods for sim-to-real transfer on physical humanoid platforms like the Unitree G1.
“The paper demonstrates that the mathematical safety constraints used to train a robot must be explicitly matched to the quality of the robot's sensors.”
Source→“We therefore deploy a lightweight Link-CBF policy zero-shot on the Unitree G1 in the real world, where it tolerates imperfect perception, succeeds on 95% of throws”
Source→“usable barrier structure depends on perceptual observability”
Source→“We theorize that the Joint-CBF gap comes from perception: the policy is asked to internalize a whole-body barrier that it cannot perceive accurately enough, and it learns worse evasions instead”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.