Watch, Infer, Coordinate: Inferring Robot Partner Constraints for Zero-Shot Coordination
The paper tests whether a robot can infer another robot’s hidden physical limits by watching how it coordinates with a teammate.
The authors introduce Watch, Infer, Coordinate, a benchmark for physically coupled manipulation tasks where one robot’s constraints may be unknown. The problem is that a demonstration shows what the constrained robot did, not everything it could have done. Their method scores possible constraints using the joint behavior of both robots, since the partner may compensate in ways that reveal the limitation. Across three settings, it improves constraint inference and zero-shot coordination, approaching an oracle given the true constraints. ArXiv · AI/CL/LG's note
The authors introduce Watch, Infer, Coordinate, a benchmark for physically coupled manipulation tasks where one robot’s constraints may be unknown. The problem is that a demonstration shows what the constrained robot did, not everything it could have done. Their method scores possible constraints using the joint behavior of both robots, since the partner may compensate in ways that reveal the limitation. Across three settings, it improves constraint inference and zero-shot coordination, approaching an oracle given the true constraints. ArXiv · AI/CL/LG's note
score 3