Task-aware motion planning in constrained environments using GMM-informed RRT planners

摘要

This paper introduces a novel integration of Task-Parameterized Gaussian Mixture Models (TP-GMM) with sampling-based motion planners, specifically RRT, to improve planning efficiency and path optimality in constrained robotic manipulation tasks. The proposed GMM-RRT and GMR-RRT planners exploit a TP-GMM trained offline on human demonstrations to generate task-adaptive sampling distributions, effectively guiding the search toward feasible and high-quality solutions. The framework is implemented in the MoveIt motion planning framework and evaluated across five simulation experiments and 30 real-world trials, focusing on Electric Vehicle (EV) battery disassembly tasks. Compared to baseline sampling-based planners, the GMM-informed planners demonstrate superior performance in key planning metrics. In the path length aspect, GMM planners yield significantly shorter trajectories, averaging 0.8 meters versus over 2 meters for baseline planners. Similarly, in path simplification time, the near-optimal nature of the generated paths reduces post-processing efforts. While planning time is higher due to TP-GMM inference and projection stages, over 90% of that time is spent outside the RRT search itself, which completes quickly due to guided sampling. Path duration also remains competitive, with GMM-informed planners closely matching RRT*. These results highlight the effectiveness of task-conditioned sampling in unstructured manipulation scenarios. The proposed method maintains 100% success rate while improving efficiency, suggesting strong potential for integration in sequential and adaptive robotic systems. Future work will focus on extending generalization to broader task parameter spaces and addressing inverse kinematics challenges. • Introducing GMM-Informed planners which combine TP-GMM with RRT-based planning. • Learned task priors guide exploration in high-dimensional configuration space. • Planning involves task-to-joint projection via Jacobian-based inverse kinematics. • A Benchmark is provided across varied scenes of scattered obstacles. • Real-world validation in EV battery disassembly scenario.