Collaborative Multiagent Closed‐Loop Motion Planning for Multimanipulator Systems

Multimanipulator robotic cells are increasingly employed to satisfy stringent cycle‐time and dexterity requirements through parallel operations and coordinated multiarm actions. The increased co‐occupancy of shared workspaces elevates self‐ and inter‐robot collision risks and complicates motion coordination. This study presents a hierarchical closed‐loop control architecture that integrates high‐level multiagent reinforcement learning (MARL) for global strategy with a low‐level motion layer based on enhanced optimized normalized dynamic movement primitives (ON‐DMP*). The ON‐DMP* model automates parameter selection, eliminates manual tuning, generating smooth and collision‐aware joint trajectories in real time. At the supervisory level, the MARL coordinator dynamically resolves contention in overlapping regions, assigns subgoals, and adapts to disturbances based on shared feedback. Experimental validation on assembly and disassembly tasks with temporal and spatial overlap demonstrates consistent avoidance of self‐ and mutual‐collision, improved coordination smoothness, and reduced task completion time compared to reinforcement learning‐based and noncooperative movement primitive baselines. The results indicate a versatile approach to scalable cooperative manipulation in shared work environments. Video Link: https://youtu.be/G‐ON6DK62YQ