Collision-Free Cooperative MegaCRN-Enhanced Model for Multi-Manipulator Motion Planning in Shared Workspaces

Abstract

Shared-workspace robotic cells increasingly require multiple manipulators to operate simultaneously for part exchange, tool changing and temporary access to confined spaces. This paper develops a MegaCRN-CM model for collision-free cooperative motion planning, which enhances MegaCRN with collision memory, kinematic graph attention, and a reservation decoder for swept-volume conflicts. Each manipulator is encoded as a joint-link-tool graph, cell-time occupancy over a short horizon is predicted, and then bounded corrections are applied to waypoints and timing before deterministic geometric verification. Six shared-workspace layouts and six-axis industrial manipulators were selected for the experiments, with 2,400 simulated production cycles and 360 hardware-in-the-loop cycles that randomly varied fixtures, payloads and handover windows. MegaCRN-CM reduced the number of near-collision events per 100 cycles from 18.7 in fixed-priority planning to 2.1 and increased the median minimum inter-link distance from 41 mm to 72 mm. The four task sets of the new model achieved an average throughput of 91.6%, exceeding that of unmodified MegaCRN by 2.2% and fixed-priority planning by 19.2%. The 95th-percentile replanning latency was 37.8 ms and still within the 40 ms industrial planning interval. According to ablation experiments, collision memory and swept-volume reservation have shown the largest safety and waiting-time improvements. The proposed model offers a feasible learning-assisted planning structure for cooperative manipulators in confined and dynamically shared workspaces.