New LLM architecture enables robots to collaborate with humans

A team of researchers from UC Santa Barbara and Florida State University has introduced a structured large-language-model (LLM) architecture designed to enable zero-shot human-robot coordination in cooperative construction tasks with hidden goals. Published on arXiv, the work presents a novel approach that decomposes the team decision problem into five key components: action-conditioned Theory-of-Mind (ToM) inference, hierarchical planning, conversation interpretation, action verification, and feedback-based replanning. This architecture was benchmarked against two baselines: an ablation without ToM inference and a multi-agent reinforcement learning (RL) policy trained offline on numerous goal pairs.

In human-participant experiments, the proposed LLM-based method required 30% fewer interaction steps and achieved 25% higher post-interaction trust ratings compared to the baselines. The researchers attribute this success to systematically decomposing the decision-making process and using LLMs as tractable surrogates for otherwise intractable inference and planning computations while retaining conventional verification for physical feasibility. This breakthrough suggests a pathway to more intuitive and efficient human-robot collaboration in real-world settings.