Partially Observable Zero-shot coordination by Predicting Intention of Partner
PIP tries to keep coordination stable when a partner drops out of view.
The paper targets zero-shot embodied coordination under partial observability, where hidden partner states make existing representations uncertain. Its method combines a Joint-view VAE for richer partner representations with belief networks that infer hidden location and behavior from interaction history. The authors test it on Burrito-PO, Overcooked-PO, and a Melting Pot substrate, plus a human evaluation in Burrito-PO. They report the highest mean performance among compared methods across all three benchmarks. Source: ArXiv · AI/CL/LG's note.
The paper targets zero-shot embodied coordination under partial observability, where hidden partner states make existing representations uncertain. Its method combines a Joint-view VAE for richer partner representations with belief networks that infer hidden location and behavior from interaction history. The authors test it on Burrito-PO, Overcooked-PO, and a Melting Pot substrate, plus a human evaluation in Burrito-PO. They report the highest mean performance among compared methods across all three benchmarks. Source: ArXiv · AI/CL/LG's note.
score 4