DexCompose: Reusing Dexterous Policies for Multi-Task Manipulation with a Single Hand

A new framework called DexCompose can reuse pretrained dexterous manipulation policies to perform multiple tasks with a single robotic hand, achieving a 77.4% average composite success rate across 16 tasks, according to a preprint posted to the arXiv repository on June 26, 2026 [1][2]. Dexterous manipulation policies can solve individual skills, but composing them to perform multiple tasks with a single hand remains challenging because adding a new task often imposes conflicting demands on overlapping fingers and contact modes [2]. The DexCompose framework addresses this by assigning explicit finger-level action ownership. It first collects successful post-task states from an initial skill and performs release tests over candidate finger masks to identify which fingers are necessary for maintaining the established state [2]. It then trains two asymmetric residual modules: a bounded residual stabilizer for task preservation, and a context-aware residual that adapts a frozen downstream policy only within the action subspace assigned to the new task [2]. The evaluation spanned 16 composite dexterous manipulation tasks covering four object-retention skills and four downstream interactions [2]. The 77.4% average composite success rate indicates that structural action ownership with dual residuals offers a direction for composing dexterous skills beyond conventional policy chaining [2]. The paper was submitted under the Computer Science > Robotics category on arXiv, an open-access repository of electronic preprints that is not peer reviewed [6]. arXiv was founded on August 14, 1991, and as of November 2024 receives about 24,000 article submissions per month [6]. The repository hosts a framework called arXivLabs, which allows community collaborators to develop and share experimental tools directly on article record pages [5]. arXivLabs projects have included the Bibliographic Explorer for citation-tree navigation and the CORE Recommender for discovering related open-access papers [4][5]. The DexCompose preprint appears alongside these Labs integrations on its abstract page [1]. The work contributes to ongoing robotics research into dexterous manipulation, where reusing pretrained policies for multi-task settings has been an open problem [2]. The authors note that destructive interference between preserving an existing manipulation outcome and executing a new one has been a central obstacle [2]. By explicitly modeling which fingers are responsible for which task, DexCompose reduces this interference without requiring retraining of the original full-hand policies [2].