Reinforcement Learning-enhanced Adaptive Task Planning for Household Embodied Agents with Large Language Models

While large language models (LLMs) emerged as a powerful tool for planning and reasoning, how to utilize them for task planning for embodied agents within complex, real-world environments like households environments, remains a significant challenge. While LLMs possess extensive commonsense knowledge, their planning performance often suffers from low success rates due to a lack of grounding in specific environmental constraints and limited executable action sets. To address this issue, this paper proposes an embodied agent task planning method. Our approach (1) utilizes the LLM as a planner that generates subgoal sequences based on linguistic representations of the environment state and the ultimate goal, and (2) employs the same LLM as a score function to evaluate and select the most probable action from a predefined set. This dataset is then used to fine-tune a base LLM via GRPO algorithm, which aligns the model’s outputs more effectively with successful planning trajectories. We evaluate our method on SayCan and LongTasks benchmark. Experimental results demonstrate that our method achieves a substantially higher task success rate, outperform baseline methods by 22.29 % in Success Rate (SR) and 19.64% in Executability (Exe).