Reproduce “Robot Learning via Human Adversarial Games” on OpenResearch

About this paper

Jiali Duan, Qian Wang, Lerrel Pinto, C.-C. Jay Kuo, Stefanos Nikolaidis

University of Southern California and Carnegie Mellon University researchers propose using human adversarial actions to enhance robot learning, demonstrating that a robot trained to grasp objects against human attempts to dislodge them achieves significantly higher grasp robustness compared to self-supervised learning, particularly for challenging objects.

Key insights

• Human adversaries possess inherent domain knowledge about object geometry and physics, enabling them to apply targeted perturbations that efficiently expose weaknesses in a robot's grasping policy.
• Framing human-robot interaction adversarially, rather than cooperatively, provides a valuable and efficient source of training data for improving robot robustness.
• The effectiveness of human adversarial training is particularly pronounced for objects with a limited number of robust grasp candidates, where targeted feedback guides the robot more efficiently.

Summary by alphaXiv.