Coupling-Aware Planner Exploration for Shared-Workspace Multi-Manipulator Motion Planning

Abstract

Shared-workspace multi-manipulator planning repeatedly faces a representation decision: exploit efficient decoupled/hybrid planning or escalate to coupled composite-space search. We cast this as coupling-aware exploration over planning representations: the system probes interaction structure before deciding whether to exploit cheap decoupled/hybrid planning or explore a more expensive coupled composite-space representation. Our framework logs interpretable interaction descriptors, routes tasks among decoupled, hybrid, and joint-space modes, and evaluates these policies on 4,000 PyBullet tasks and 32,000 planner runs. Coupling-aware routing improves paired solve outcomes over random and pure joint-space routing, while matching the best fixed-mode solve rate in the selection ablation. The main contribution is a regime-level account of escalation under interaction uncertainty: coupling descriptors indicate when hybrid exploitation is sufficient, when deeper coupled search is warranted, and where selector overhead remains.