MUCO: A Mani-Unit-Structured Manipulation Framework for Compositional Long-Horizon Robot Tasks

This research proposes MUCO, a framework for long-horizon robot manipulation tasks that uses reusable manipulation units (MUs) defined by local goals, initiation conditions, and success/failure criteria. MUC, through MU-Centric Orchestration, constructs MU sequences, matches them to compatible policy models, and enables retry or fallback without re-executing verified steps. Evaluated on RLBench, RoboTwin 2.0, and real-world tasks, MUCO achieves higher success rates, better failure containment, and reduced error propagation compared to baselines.