Partially Observable Zero-shot coordination by Predicting Intention of Partner
PIP improves zero-shot multi-agent coordination when partners are intermittently out of view.
The paper proposes Predicting Intention of Partner (PIP) for zero-shot coordination when a partner is intermittently occluded. A Joint-view VAE distills training-time evidence from both agents’ observations into a representation usable from local views, while Partner-state Belief networks infer hidden location and behavioral tendencies. PIP records the highest mean performance among compared methods on Burrito-PO, Overcooked-PO, and a Melting Pot substrate. Human evaluation in Burrito-PO and diagnostic analyses support coordination with unseen partners and the value of both components.
- PIP targets coordination when a partner is intermittently out of view.
- A Joint-view VAE builds a partner representation from local observations.
- Belief networks infer hidden location and behavioral tendencies.
- Highest mean score on Burrito-PO, Overcooked-PO, and a Melting Pot substrate.
- Human evaluation in Burrito-PO supports coordination with unseen partners.
From the original article
1 min read at arxiv.orgZero-shot coordination in embodied settings requires acting while the partner is intermittently out of view, leaving existing methods with ambiguous partner representations and uncertainty over hidden partner states. We propose Predicting Intention of Partner (PIP) to jointly address these challenges.
By Jinnyeong Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.