Zero-shot coordination in embodied environments must cope with partners that are intermittently out of view, which makes partner representations ambiguous and hidden states uncertain. To tackle both issues we introduce Predicting Intention of Partner (PIP). PIP employs a Joint‑view VAE that aggregates the full set of local observations from both agents during training, distilling a partner representation that can later be recovered from a single agent’s local view alone. A Partner‑state Belief network then uses the ego‑agent’s interaction history to infer the partner’s hidden location and behavioral tendencies. We benchmark PIP on Burrito‑PO, Overcooked‑PO, and a Melting Pot substrate, and also conduct a human study in Burrito‑PO. Across all three tasks PIP achieves the highest mean performance among compared methods. Human evaluation and diagnostic analyses further confirm effective coordination with unseen partners under occlusion and the complementary roles of the two components.
Review