Robots that Learn to Safely Influence via Prediction-Informed Reach-Avoid Dynamic Games

• DOI: 10.48550/arxiv.2409.12153
• Publication Date: 2024-09-18
• Authors: Ravi Pandya; Changliu Liu; Andrea Bajcsy
• Keywords: FOS: Computer and information sciences; Computer Science - Robotics; Robotics (cs.RO)

Robots can influence people to accomplish their tasks more efficiently: autonomous cars inch forward at an intersection pass through, and tabletop manipulators go for object on the table first. However, a robot's ability also compromise safety of nearby if naively executed. In this work, we pose solve novel robust reach-avoid dynamic game which enables robots be maximally influential, but only when backup control exists. On human side, model human's behavior as goal-driven conditioned plan, enabling us capture influence. robot in joint physical belief space, reason about how its uncertainty will evolve over time. We instantiate our method, called SLIDE (Safely Leveraging Influence Dynamic Environments), high-dimensional (39-D) simulated human-robot collaborative manipulation task solved via offline game-theoretic reinforcement learning. compare approach baseline that treats worst-case adversary, controller does not explicitly influence, energy-function-based shield. find consistently leverage it has is safe do so, ultimately allowing less conservative while still ensuring high rate during execution.