Multi-Agent Probabilistic Grounding Enables Robots to Resolve Ambiguous Language Commands via Consensus

arXiv·low signal

arXiv paper (2603.19166) introduces a multi-agent probabilistic grounding framework for vision-language navigation where multiple agents collaboratively convert ambiguous natural language goals into physically-actionable, grounded decisions. The framework addresses the core alignment failure mode in embodied agents — where a single agent's language-to-action mapping produces physically incorrect or unsafe behavior — by using probabilistic consensus across agents to reduce grounding error. This pattern is directly applicable to any multi-agent system where task specifications are underspecified natural language.