DexCompose: Reusing Dexterous Policies for Multi-Task Manipulation via Finger-Level Action Ownership

Forum topic · 小凯 · 2026-06-30

Summary

DexCompose is a role-aware residual composition framework that reuses pretrained dexterous manipulation policies for multi-task control with a single hand. Composing skills is difficult because a new task often imposes conflicting demands on overlapping fingers and contact modes, causing destructive interference between maintaining an existing manipulation outcome and executing a new one. DexCompose addresses this with explicit finger-level action ownership: given two pretrained full-hand policies, it collects successful post-task states from the first skill and performs release tests over candidate finger masks to determine which fingers are necessary to maintain the established skill state. It then trains two asymmetric residual modules—a bounded residual stabilizer for task preservation and a context-aware residual that adaptively freezes the downstream policy within the action subspace assigned to the new task. Evaluated on 16 composite dexterous manipulation tasks spanning four object-retention skills and four downstream interactions, DexCompose achieves 77.4% average composite success rate, suggesting structured action ownership plus a dual-residual design is a promising direction for dexterous skill composition beyond conventional policy chaining.

Overview

Field: Computer Vision / Robotics Authors: Dihong Huang, Zhenyu Wei, Zhuxiu Xu Published: 2026-06-26 arXiv: 2606.28323

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.

The authors propose DexCompose, a role-aware residual composition framework that reuses pretrained dexterous policies for multi-task manipulation through explicit finger-level action ownership.

Method

Given two pretrained full-hand policies, DexCompose works in two stages:

1. Finger ownership identification: Collect successful post-task states from the first skill and perform release tests over candidate finger masks to identify which fingers are necessary for maintaining the established skill state. 2. Dual residual training: Train two asymmetric residual modules:

A bounded residual stabilizer for task preservation.
A context-aware residual that adaptively freezes the downstream policy only within the action subspace assigned to the new task.

Results

The framework was evaluated on 16 composite dexterous manipulation tasks, covering four object-retention skills and four downstream interactions. DexCompose achieves a 77.4% average composite success rate, indicating that structured action ownership combined with a dual-residual design offers a promising direction for dexterous skill composition beyond conventional policy chaining.

---

*Auto-collected on 2026-06-30.*