Communicative PARTNR: Natural Language Communication under Partial Observability in Human–Robot Collaboration

HoBeomJeon, Hyungmin Kim, DohyungKim, Minsu Jang, Jaehong Kim

Keywords: Cooperative Embodied Agent, Multi Agent Coordination, Natural Language Communication

TL;DR: Communicative PARTNR shows that uncertain system feedback can improve task success in range-limited, heterogeneous human–robot collaboration, motivating future work on advanced dialogue processing.

Abstract: Natural language communication is a key mechanism for coordination in human–robot collaboration under partial observability, where agents possess only local views of the environment. While prior work often assumes fully reliable or unconstrained message channels, real-world settings impose delivery uncertainty and range limitations that complicate when and how agents should communicate. We present Communicative PARTNR, an extension of the PARTNR benchmark that introduces a range-limited channel and four levels of system feedback information: Opaque (no confirmation), Binary (success/failure), Causal (failure reason), and Traceable (failure reason with partner state). Using LLM-based embodied agents in decentralized household tasks, we find that minimal and unclear feedback (Opaque) yields higher task success and completion rates than richer alternatives, despite generating fewer and shorter dialogues. Analysis reveal that excessive detail can divert agent reasoning from task execution, whereas concise system feedback maintain focus and coordination efficiency. These results underscore the importance of designing dialogue strategies and context representations that enable agents to exploit communication outcome information effectively without incurring unnecessary cognitive or temporal overhead. Code is available in https://github.com/HoBeom/Communicative-PARTNR

Submission Number: 9