Whole-Body Contact-Rich Compliant Motion Planning Based on Morphology Representation

Robots operating in cluttered environments require whole-body contact-rich motion planning to safely interact with surrounding objects. However, model-based methods struggle with unknown interaction dynamics, while learning-based methods often depend on task-specific demonstrations or rewards. This paper proposes a physics-model-guided motion planning framework that learns task-agnostic whole-body interactions through material-point morphology representation and interaction prediction supervision. The representation encodes distributed body motion and tactile states with consistent body correspondence, accommodating time-varying contact configurations. A tactile-driven interaction planner generates physically informed motion references to supervise MorphoPlanner, an end-to-end network that predicts future whole-body motion morphologies from task goals and current interaction states without task-specific demonstrations or reward design. A dataset of 749 K tactile interaction samples is constructed for training and evaluation. Simulation and real-world dual-arm experiments demonstrate compliant interactions and improved planning performance over representative baselines, achieving 95.5–100% success rates in dense cluttered scenarios. These results demonstrate the potential of morphology-based physical supervision for whole-body interaction planning in unknown environments.