CLAMP: Clearance Learning Model-Driven Adaptive Motion Planning for Dual-Arm Robots

Motion planning for dual-arm robots in high-dimensional configuration spaces encounters challenges in evading self-collisions and complex, dynamic environments. This article proposes an adaptive motion planning algorithm (CLAMP) empowered by a clearance learning model-driven collision estimation module (CLM-CE). CLM-CE decouples self-collision and obstacle detection into two neural submodels: the pairwise link clearance network for self-collision prediction between dual-arm links, and the clearance field network for voxel-based environmental obstacle perception, thereby enhancing both accuracy and environmental adaptability. CLAMP leverages the CLM-CE module to drive adaptive batch expansion and shortcut filtering, accelerating path discovery and enhancing path quality. To handle uncertain and dynamic scenes, an iterative path replanning mechanism is introduced, combining geometric validation with local trajectory refinement to guarantee safe execution. Extensive experiments across diverse scenarios demonstrate that CLAMP outperforms state-of-the-art baselines by up to 83.8% in planning time and 39.5% in path length, while maintaining robust adaptability without model retraining.