Multi-agent path planning and reinforcement learning for wire harness handling

Wire harness manufacturing remains one of the most labor-intensive parts of automotive production, with the “form and tape” assembly process constituting the key bottleneck. Emerging industrial concepts shift automation from single-wire routing to module-based forming. This leads to a multi-agent planning and control problem in which moving connector housing fixtures are strongly coupled by the partially formed harness. The contact-rich and highly variable behavior of the wires as linear deformable objects is difficult to model analytically. We investigate two planning and control paradigms within a MuJoCo-based simulation environment with a Gymnasium-compatible interface. First, we use multi-agent path planning ideas on continuous, cable-coupled motion via prioritized planning with heuristic guidance and reactive constraint checking, yielding a 78% overall success rate for the best planner configuration. Second, we apply reinforcement learning, both standalone and as a residual on top of a geometric base policy, comparing SAC, PPO, and TRPO. Surprisingly, the best results are achieved by plain end-to-end SAC, which reaches a mean overall success rate of 98% (range: 95%–99%) and solves nearly 114 of 120 complete target permutation, against 65 for the best planner, whereas residual RL, despite its greater stability on single transitions, fails to generalize to the full sequence of target configurations. The results indicate a central trade-off between rule-based algorithmic planning and learned control in wire harness handling. The simulation environment is publicly available at: https://github.com/ludwigstr/WireHarness-MultiAgent-RL .