- The paper proposes a novel compositional framework leveraging finger-level policy reuse that improves composite success by up to 15.8% over baselines.
- It uses dynamic finger attribution via release tests and dual residual stabilization to separate grasp maintenance from downstream manipulation tasks.
- Empirical results on 16 composite tasks, with retention metrics of 0.811 and 0.893, confirm the strategy's scalability and performance in dexterous robotics.
Compositional Dexterous Manipulation via Policy Reuse in High-DoF Robotic Hands
Introduction and Motivation
The problem of composing dexterous manipulation policies to achieve multi-task behaviors with a single hand presents substantial challenges due to the shared and highly coupled action space of the hand. Traditional sequential chaining of pre-trained policies fails in this setting as the full-hand control required by downstream tasks interferes with grasp stabilization necessary for object retention, a phenomenon described as destructive interference. Existing alternatives—such as retraining for every composite or hard-wired DoF allocation—are unscalable, particularly given the vast space of possible skill combinations afforded by large datasets and libraries of single-task policies.
DexCompose introduces a compositional policy framework that treats policy reuse as a resource-allocation problem in the embodiment: it leverages finger-level redundancy by dynamically assigning fingers to policy subspaces and enforces this allocation through dual-residual stabilization. This approach provides an explicit mechanism for structured action ownership, enabling robust sequential or concurrent multi-skill hand manipulation without retraining entire policy sets.
Technical Approach
DexCompose comprises two main algorithmic components: finger attribution and dual residual stabilization.
Finger Attribution via Release Tests
Given two frozen, pretrained full-hand diffusion policies, the method first identifies, post-hoc, which fingers are essential for maintaining the initial manipulation result (e.g., a grasp). Using a systematically constructed set of candidate binary finger masks, the system executes controlled release tests from a library of successful Task A states. The essential metric is the object retention rate under each mask, balanced against how many fingers are freed for Task B and the cleanliness of the release (minimal residual contact from released fingers). The allocation that optimally trades off retention, action space for Task B, and anatomical appropriateness is selected—this assignment is often conducted using an LLM agent that contextually reasons beyond pure retention statistics to optimize downstream manipulation sufficiency.
Dual Residual Stabilization
After finger allocation, the action space is explicitly partitioned. The subset retained by Task A is stabilized by a bounded joint-space residual policy trained to reject perturbations from Task B, thereby isolating grasp maintenance in the restricted subspace. Conversely, Task B controls the released finger joints and the wrist DoFs, with its own context-aware residual module correcting for the constraints induced by an already-engaged grasp.
Both residuals are lightweight (small PPO-trained policies over a restricted subspace) and initialized near zero, ensuring they minimally disrupt the base behaviors. The overall composed action is a mask-weighted sum of the preserved-finger residual-stabilized action for Task A and the residual-corrected output of Task B for the released subspace.
Experimental Results
DexCompose is evaluated on 16 composite manipulation tasks in simulation, using the 24-DoF Shadow Hand across a matrix of four object-retention skills (e.g., GraspBall, PickStick) and four downstream manipulation skills (e.g., OpenDoor, PushButton). Each composite task requires maintaining the outcome of Task A (retention/stability) while successfully performing Task B (interaction/actuation).
Major findings:
- Composite Success Rate: DexCompose achieves a mean composite success rate of 77.4%, which is a 15.8% absolute improvement over the strongest baseline (residual learning without action masking) and a 74.3% improvement over direct sequential policy chaining.
- Policy Preservation: Both A-side (retention) and B-side (downstream skill) preservation metrics are maximized by DexCompose (0.811 and 0.893, respectively), indicating minimal destructive interference.
- Ablations: The Task A residual stabilizer proves essential for outcome preservation. Removing finger allocation or masking degrades performance significantly, confirming the necessity of explicit subspace partitioning. A single unmasked residual (as in classical residual RL) is far less robust due to unchecked interference.
- Mask Selection: LLM-guided finger masking outperforms heuristic retention-maximizing selection by an average of 6.5%, validating the importance of task-aware, semantic reasoning in redundancy allocation.
- Baseline Comparisons: All conventional approaches, including decomposed action subspaces with fixed allocation and unstructured residual blending, underperform, especially on interaction-intensive compositions.
Implications and Future Directions
DexCompose establishes a scalable paradigm for composition of dexterous skills by leveraging hand embodiment redundancy and enforcing structural action-ownership constraints. This directly enables practical reuse of large libraries of learned single-skill policies in compound tasks, without the combinatorial explosion of retrained controllers.
Theoretically, the work advances the formalization of compositional control in high-DoF, resource-sharing agents, demonstrating that explicit action subspace partitioning is critical for mitigating cross-task interference. It also demonstrates the utility of post-hoc resource attribution, a direction that can generalize to other embodied agents (e.g., legged locomotion with concurrent manipulation).
Practically, this permits rapid prototyping or reconfiguration of task graphs in robotic platforms with dexterous effectors, reducing data and engineer-ing overhead for integrating new downstream tasks. The LLM-based attribution hints at future avenues for combining statistical metrics and semantic knowledge for compositional RL.
Limitations include the restriction to sequential two-task combinations; extensions to multi-stage task graphs or continuous task blending represent ongoing challenges. Further research may explore recursive or hierarchical action subspace decompositions and scaling to bimanual settings.
Conclusion
DexCompose provides a rigorous framework for composing dexterous manipulation behaviors through structured policy reuse, explicit resource allocation, and subspace-specific stabilization. By formulating the action space allocation and stabilization problems as first-class components, DexCompose delivers state-of-the-art robustness in multi-task hand manipulation, highlighting concrete methods to sidestep interference and maximize the utility of learned behaviors for flexible embodied AI. The approach yields clear improvements in success rates, preservation metrics, and practical scalability, delineating a strong foundation for next-generation compositional control in dexterous robotics.