Two robots do not automatically dismantle a battery twice as quickly as one. They need to agree on which parts are available, respect the order imposed by the assembly, and avoid obstructing each other; this paper investigates that coordination problem in a ROS/Gazebo battery-disassembly simulation.
Why Disassembly Needs a Task Planner
Disassembly changes the scene after every successful action. Removing a cover exposes components; taking a module changes what remains; a part selected by one arm must not simultaneously remain a target for the other. A planner must therefore reason about both prerequisites and the current scene rather than repeatedly selecting the nearest visible object.
The system combines logical task selection with motion planning. The task layer chooses an eligible component and assigns an action, while a motion planner searches for a feasible robot trajectory. These are complementary responsibilities: correct sequencing does not imply collision-free motion, and a collision-free path to the wrong component does not complete the intended task.
From Images to Coordinated Actions
Cameras attached to the end-effectors provide images for component detection and segmentation. Estimated object positions must be transformed from camera coordinates into the frame used for robot planning. Object identity, hierarchy and distance then contribute to selecting what should be handled next.
Hierarchy prevents an attractive nearby target from being selected before its prerequisites are satisfied. Distance helps choose among eligible targets, while a shared planning representation accounts for the two arms occupying the same workspace. Grasping and placement update the task state so that later decisions use the changed scene.
This design illustrates why task allocation cannot be reduced to independent arm programs. A robot’s decision consumes an object and may occupy space needed by its neighbour. Coordination requires consistent object state as well as geometric information about both manipulators.
The Simulation and Its Assumptions
The evaluation uses two UR10 models with camera-equipped vacuum grippers and a simplified Nissan e-NV200 battery pack in Gazebo. Bolts are assumed already unbolted, and some physical properties were disabled to avoid unstable simulated interactions. The experiment consequently studies retrieval and coordination under those assumptions, not every contact-rich operation in actual disassembly.
Those simplifications matter to interpretation. A model can validate sequencing and trajectory generation without accurately representing a seized fastener, deforming cable or uncertain suction seal. A realistic picture of the workcell does not, on its own, make its contact dynamics physically validated.
What the Planner Comparison Shows
| Motion planner | Reported simulated disassembly time |
|---|---|
| RRT | 543.06 s |
| RRTConnect | 541.89 s |
| RRTStar | 547.27 s |
These sampling-based planners search robot configuration space differently, but achieved similar completion times in the reported setup. The fastest-to-slowest spread is 5.38 seconds, approximately 0.99% of the fastest result. That observation supports compatibility with all three planners here; it does not establish statistical equivalence across arbitrary layouts or random seeds.
The paper also examines trajectories and path length. A shorter geometric path need not yield proportionally lower total cycle time when grasping, placement, synchronisation and execution speeds contribute to the duration. Likewise, a second arm helps only where prerequisites and shared-space constraints permit useful concurrency.
Interpretation and Limits
The main value is a complete perception–task–motion example in which the two robots respond to the evolving disassembly scene. It provides a basis for investigating alternative scheduling or perception methods while retaining explicit interfaces between them.
Transferring the result to hardware would require validation of real perception, grasping, contact behaviour and fault handling. Reproducing the numerical experiment requires the original scene, models, detector configuration, planning budgets and timing definitions; the paper lists data as available on request. Simulated cycle times should remain labelled as such.
Testing and Real-World Use
The coordination approach could be tested on repeated simulated layouts while logging prerequisite violations, duplicate assignments, collisions and completion time for identical planner settings. It could help design multi-arm recycling or assembly cells where robots must share objects and workspace without conflicting actions.