Partially Observable Zero-shot coordination by Predicting Intention of Partner

TL;DR

The paper proposes Predicting Intention of Partner (PIP) to address zero-shot coordination challenges in embodied settings. PIP uses a Joint-view VAE and Partner-state Belief networks to infer partner representations and hidden states. It achieves the highest mean performance across three benchmarks.

Why It Matters

PIP enables effective coordination with unseen partners in partially observable environments, which is crucial for real-world applications like human-robot collaboration. This contributes to the development of more robust and adaptive AI systems. It also has implications for multi-agent reinforcement learning and human-AI interaction.

Key Findings

• PIP outperforms existing methods in zero-shot coordination
• Joint-view VAE effectively distills partner representations
• Partner-state Belief networks accurately infer hidden partner states