LLM-driven symbolic planning and hierarchical imitation learning for long-horizon deformable object assembly
Jiaming Qi, Liang Lu, Fangyuan Wang, Hoi-Yin Lee, David Navarro-Alarcón, Zeqing Zhang, Peng Zhou
Abstract
Long-horizon assembly tasks involving deformable objects pose substantial challenges for autonomous robots, stemming from infinite-dimensional state spaces, complex sequential dependencies, and high variability in real-world conditions. In this work, we propose a novel and robust framework that tightly integrates Large Language Model (LLM)-driven symbolic planning with hierarchical imitation learning to enable reliable and generalizable solutions for deformable object assembly. Our approach leverages the advanced reasoning capabilities of LLMs to translate natural language task instructions into structured symbolic task plans. This decomposition is initiated by a visual-language model (VLM) that generates descriptive subgoals from key visual frames of a human demonstration. Each subgoal is then automatically grounded in the robot’s perception via a VLM query mechanism, ensuring precise and task-relevant state estimation. For execution, a 3D diffusion policy (DP3) conditioned on visual input and symbolic subgoals generates smooth, low-level action trajectories, bridging the gap between high-level symbolic reasoning and dexterous manipulation. We validate our hierarchical framework on a real-world round belt drive assembly benchmark, demonstrating significant improvements in success rates, error recovery, and generalization across diverse and perturbed initial conditions, compared to existing approaches. Our results highlight the potential of integrating LLM-based symbolic abstraction, targeted state querying, and diffusion-based visuomotor control for robust, autonomous assembly of deformable objects in unstructured environments.