Robotics

PEEL Races Parts Out of Assemblies With Parallel Motion Planning

Robb Harlan 5 min read

Disassembly sequence planning often builds a huge precedence graph first, then validates motions. That combinatorial front-end does not age well past a few dozen parts.

PEEL (Parallel Extraction for Long-Horizon Disassembly), arXiv:2608.08773 (Aug 9, 2026) from Servet B. Bayraktar, Andreas Orthey, Zachary Kingston, and Marc Toussaint, flips the order: geometric removability is the sequence. Batches of single-part planners race; the first to find a collision-free escape wins, competitors die, and the reduced assembly goes again.

Fetch manipulator disassembling a pair of pliers
Fetch mobile manipulator mid-disassembly of pliers. Source: arXiv:2608.08773.

Planner stack

Each single-object query is MAB-RRT: a multi-arm bandit switches among uniform, scale-invariant, and PCA-directional samplers to find narrow escape corridors (think bolts in tight bores). Scale is estimated in a burn-in sphere of free space; PCA then biases along the corridor.

PEEL pops up to B parts from a shuffled queue, launches concurrent MAB-RRT processes with shared time budget T, and either commits a winner or doubles T after a full failed sweep. Pure geometry: rigid bodies, zero penetration, no friction model.

Project page with animations and code pointers: peel-disassembly.surge.sh.

Results

  • MAB-RRT alone: 100% success on 76 collision-free Automate assemblies (10 random rotations each → 760 trials), under strict collision checking, outperforming AssembleThemAll baselines in the paper’s comparison.
  • PEEL on four multi-part assemblies with 10–17 removable parts, with full robot grounding (grasp → extraction → regrasp → place) on a Fetch in simulation.
  • Free-state detection uses translational/rotational mobility rank along the object path so the arm does not follow the entire overshot RRT tree into free space.
Fetch placing disassembled parts on a table
Parts placed after sequenced extraction. Source: arXiv:2608.08773.

A Human’s Take

I like treating “can I pull this part out without hitting anything” as the definition of the next step instead of a post-check on a symbolic plan. That will not replace physics for gravity-sensitive subassemblies, and the authors know it — pure geometry is a candidate generator. For recycling cells that already have CAD of bolted machinery, a 100% single-part extraction rate on a filtered Automate set is a useful bar. Show me a physical Fetch or industrial arm on one of those 17-part assemblies next.

Sources