Trust-Aware Human–Robot Collaborative Decision-Making With LLM-Generated Data

doi:10.1109/tcds.2025.3628551

Trust-Aware Human–Robot Collaborative Decision-Making With LLM-Generated Data

2025-11-03

0

PRE

AI

DOI：10.1109/tcds.2025.3628551

Original

Original request for help

Save

Abstract

En 中文

Achieving superior human–robot collaboration requires the robot to consider the dynamic trust of the human partner in decision-making. Existing research often oversimplifies tasks due to limited real human–robot collaboration data. Therefore, overcoming the scarcity of real human–robot collaboration data is essential. To address this, we utilize a large language model (LLM) to simulate various types of decision-makers and collaborate with robots in sequential tasks to acquire simulated datasets, including trust and action data. We also propose a two-stage training framework to improve the robot’s decision-making ability. In the first stage, we pretrain the trust prediction model and the decision-making model based on the simulation dataset. In the second stage, these models are fine-tuned with a small amount of real data. Our framework effectively leverages prior knowledge from simulated data while bridging the gap between LLM-generated and real human distributions. To validate our method and framework, we built a 3-D virtual environment and designed human–robot collaboration sequential tasks. We conduct experimental validation involving 39 participants. In the experiment, our trust prediction model achieves an RMSE of 0.0603 in trust prediction tests. Our trust-aware decision-making model demonstrates a 120.9% increase in reward compared with the perception-based method.

Keywords:

Decision-making

human–robot collaboration

large language model (LLM)

trust prediction

Trust-Aware Human–Robot Collaborative Decision-Making With LLM-Generated Data

Abstract

Journal

Organization