Task Planning for Dual-Arm Robot Empowered by Large Language Model

Autonomous decision-making capabilities are a key indicator of a robot's level of intelligence, requiring it to have scene understanding and reasoning abilities. Leveraging the powerful general intelligence of large language models (LLMs) for robot task planning has become an important approach in recent years. However, current research encounters several issues: (1) The output of LLM exhibits inherent randomness, making it challenging to ensure system stability and safety when directly executing LLM-generated code. (2) Existing frameworks are often designed for single-robot applications, and their direct application to multi-robot scenarios can lead to improper task allocation, task conflicts, and other issues. In this study, we propose DRTP-LLM, a framework that leverages LLM for reasoning and decision-making in dual-arm robotic tasks while ensuring system safety. This framework includes the prompt mechanism, a library of primitives, and the primitive sequence extractor, among other components. We deployed this method on a real dual-arm robot system and validated it in tabletop scenarios, evaluating the results with specific metrics. Compared to directly executing LLM-generated code, our method demonstrates a higher success rate. Additionally, the optimizations for dual-arm robot scenarios lead to more reasonable and efficient task allocations.