An LLM-enabled human demonstration-assisted hybrid robot skill synthesis approach for human-robot collaborative assembly

Effective human-robot collaborative assembly (HRCA) demands robots with advanced skill learning and communication capabilities. To address this challenge, this paper proposes a large language model (LLM)-enabled, human demonstration-assisted hybrid robot skill synthesis approach, facilitated via a mixed reality (MR) interface. Our key innovation lies in fine-tuning LLMs to directly translate human language instructions into reward functions, which guide a deep reinforcement learning (DRL) module to autonomously generate robot executable actions. Furthermore, human demonstrations are intuitively tracked via MR, enabling a more adaptive and efficient hybrid skill learning. Finally, the effectiveness of the proposed approach has been demonstrated through multiple HRCA tasks.