DexCompose: Reusing Dexterous Policies for Multi-Task Manipulation with a Single Hand
2026-06-26 • Robotics
RoboticsArtificial IntelligenceComputer Vision and Pattern RecognitionMachine Learning
AI summaryⓘ
The authors address the challenge of combining different skilled hand movements so one hand can do multiple tasks without messing up previous skills. They created DexCompose, a method that figures out which fingers are needed to keep the first task's results and then trains special modules to keep that skill stable while adapting only the necessary fingers for the new task. Testing this on many tasks showed that their method works well and better manages multi-task hand control than simple policy combinations. Essentially, they help a robot hand share work between fingers more smartly when doing several tasks.
dexterous manipulationpolicy compositionresidual learningmulti-task learningrobotic hand controlfinger ownershipaction subspaceskill preservationcomposite tasksreinforcement learning
Authors
Dihong Huang, Zhenyu Wei, Zhuxiu Xu, Yunchao Yao, Sikai Li, Mingyu Ding
Abstract
Dexterous manipulation policies can solve individual skills, but composing them to perform multiple tasks with a single hand remains challenging. Adding a new task on top of an existing manipulation skill often imposes conflicting demands on overlapping fingers and contact modes, causing destructive interference between preserving an existing manipulation outcome and executing a new one. We propose DexCompose, a role-aware residual composition framework that reuses pretrained dexterous policies for multi-task manipulation through explicit finger-level action ownership. Given two pretrained full-hand policies, DexCompose first collects successful post-task states from the first skill and performs release tests over candidate finger masks to identify which fingers are necessary for maintaining the established skill state. It then trains two asymmetric residual modules: a bounded residual stabilizer for task preservation, and a context-aware residual that adapts the frozen downstream policy only within the action subspace assigned to the new task. We evaluate the framework on 16 composite dexterous manipulation tasks spanning four object-retention skills and four downstream interactions. DexCompose achieves a 77.4% average composite success rate, demonstrating that structural action ownership with dual residuals offers a promising direction for composing dexterous skills beyond conventional policy chaining.