An In-Context Learning Approach for LLM-Enabled Multi-Robot Task Allocation

In-context learning for large language models (LLMs) is employed to design a novel hybrid optimization framework for multi-robot task allocation. Results show that this approach can improve existing approaches, such as the market based scheme or the speedup and slow-down (SUSD) scheme by leveraging the contextual understanding ability of a LLM. Incorporating LLM removes the need for detailed information regarding the utility function for multi-robot task allocation. LLM helps the algorithm to escape local optima without relying on any human feedback. However, in-context learning is necessary to allow the LLM to agree with the underlying optimization algorithms. This paper proposes an method to instigate in-context learning for LLMs and demonstrate convergence to consensus between an LLM and the underlying optimization algorithms. The proposed method reduces the need for expert intervention and enhances the applicability and scalability for complex task allocation problems. Experimental results verify that the hybrid approach can improve the effectiveness of task allocations compared with recent SUSD-only methods.