DexGrasp-Zero: Rethinking Cross-Embodiment Transfer Through Morphology-Aligned Graph Representations

Introduction

The proliferation of dexterous robotic hands, from the Shadow Dexterous Hand to the LEAP Hand and Inspire Robotics platforms, presents both an opportunity and a challenge for the robotics community. While hardware diversity enables application-specific optimization, it creates a fragmentation problem: policies trained on one hand architecture rarely transfer to others without extensive retraining or hand-specific engineering. The paper "DexGrasp-Zero: A Morphology-Aligned Policy for Zero-Shot Cross-Embodiment Dexterous Grasping" by Wu et al. addresses this bottleneck through a fundamental rethinking of how we represent robotic morphologies in learning frameworks.

The core contribution lies in abandoning the retargeting paradigm that has dominated cross-embodiment research. Rather than learning intermediate motion targets that require hand-specific retargeting models (which often violate kinematic constraints), DexGrasp-Zero learns a universal policy that maps morphology-aligned graph representations directly to physical actions. This architectural shift enables zero-shot transfer to unseen hand morphologies with an 85% success rate in simulation, representing a 59.5% improvement over prior state-of-the-art methods.

The Retargeting Problem and Graph-Based Solutions

Existing approaches to cross-embodiment dexterous grasping typically rely on intermediate motion representations. These methods predict target poses or trajectories in a canonical space, then retarget these targets to specific hand morphologies through optimization or learned models. However, as Wu et al. identify, this pipeline introduces a critical failure mode: intermediate targets frequently violate the physical constraints of the target embodiment, such as joint limits or collision geometry. When a policy trained on a 20-DoF Shadow Hand attempts to control a 16-DoF Allegro Hand through retargeting, the kinematic mismatch creates infeasible configurations that degrade performance.

DexGrasp-Zero resolves this by operating directly in the physical action space of each hand through a morphology-aligned graph representation. The key insight is that despite varying degrees of freedom and link geometries, dexterous hands share anatomical regularities. fingers consist of linked segments connected by revolute joints; thumbs oppose fingers with similar kinematic structure regardless of specific hardware implementation. By mapping each hand to a graph where nodes represent anatomically grounded keypoints (metacarpals, proximal joints, distal tips) and edges represent kinematic chains, the policy learns structural invariances rather than embodiment-specific patterns.

This graph representation is augmented with tri-axial orthogonal motion primitives, grounded in biomechanical principles. Rather than learning raw joint commands, the policy outputs actions in a semantic space representing flexion/extension, abduction/adduction, and opposition movements. This biomechanical abstraction aligns control semantics across morphologies; a "flexion" command translates appropriately whether controlling a Shadow Hand's sophisticated tendon-driven finger or a LEAP Hand's direct-drive architecture.

MAGCN and Physical Property Injection

The policy architecture, termed the Morphology-Aligned Graph Convolutional Network (MAGCN), extends standard graph convolutional networks with a critical adaptation for robotics: the Physical Property Injection mechanism. While graph neural networks excel at learning relational patterns, they typically treat nodes as generic feature vectors. MAGCN explicitly incorporates embodiment-specific physical constraints derived from URDF (Unified Robot Description Format) specifications.

During forward passes, MAGCN injects link lengths, joint limits, and actuation constraints directly into the graph node features. This enables adaptive compensation during grasp execution; the policy learns to modulate grip force and finger positioning based on the specific torque limits and link geometries of the current hand. For instance, when controlling a hand with shorter distal phalanges, the policy automatically adjusts contact point predictions to ensure stable force closure, a behavior that emerges from the physical property conditioning rather than explicit programming.

The training protocol jointly optimizes across four heterogeneous hands: the Shadow Dexterous Hand (20 DoF, tendon-driven), the Allegro Hand (16 DoF, direct-drive), the Schunk SVH (20 DoF, anthropomorphic), and the Ability Hand (varied configurations). This multi-embodiment training forces the policy to extract universal grasping principles while the physical property injection handles embodiment-specific constraints. Notably, the policy achieves this without hand-specific retargeting modules; a fixed mapping function converts the motion primitive outputs to joint commands for each hand, eliminating trainable components that could introduce errors during transfer.

Zero-Shot Transfer and Sim-to-Real Validation

The experimental validation addresses both simulation and physical deployment. On the YCB dataset, the policy achieves 85% zero-shot success when transferred to unseen hands (specifically the LEAP and Inspire platforms), compared to approximately 53% for the best competing method. This gap highlights the cost of retargeting errors in previous approaches; when policies must translate through intermediate representations, compound errors accumulate, particularly for out-of-distribution morphologies.

Real-world deployment on three physical platforms, LEAP, Inspire, and Revo2 hands, demonstrates an 82% average success rate on unseen objects. This sim-to-real transfer is notable given that the policy never trained on these specific hardware platforms. The Physical Property Injection mechanism proves crucial here; by respecting the actual torque limits and kinematic constraints of the deployment hardware, the policy avoids the jerky, unstable behaviors that often plague transferred policies.

The choice of evaluation objects is significant. Rather than testing only on geometric primitives, the experiments include challenging YCB objects with varying friction coefficients, mass distributions, and surface geometries. The policy's ability to generalize across both novel hands and novel objects suggests it captures fundamental grasp affordances rather than memorizing specific hand-object configurations.

Critical Analysis and Limitations

While DexGrasp-Zero represents a significant advance, several limitations warrant discussion. The method assumes hands share sufficient anatomical similarity to map to a common graph structure. Extremely divergent manipulators, such as three-fingered industrial grippers or soft pneumatic hands without discrete joints, may not align with the current morphology graph without substantial modification. The framework essentially assumes a "hand-like" topology with opposing thumbs and articulated fingers.

Furthermore, the current evaluation focuses exclusively on grasping rather than general in-hand manipulation. Cross-embodiment transfer for tasks requiring finger gaiting, pivoting, or tool use presents additional challenges; these tasks require not just static force closure but dynamic repositioning where embodiment-specific dynamics become more critical. Whether the Physical Property Injection mechanism suffices for high-frequency manipulation tasks remains an open question.

The reliance on accurate URDF specifications introduces a practical constraint. While simulation environments provide perfect URDF data, real-world hands often exhibit calibration errors, gear backlash, or tendon stretch not captured in the nominal models. The paper does not explicitly address robustness to such model mismatches, though the high real-world success rates suggest some inherent tolerance.

Implications for General-Purpose Robotics

DexGrasp-Zero suggests a path toward truly general-purpose robotic systems where software and hardware evolve independently. Currently, new hand hardware requires either new control policies or careful retargeting engineering. By learning universal controllers that adapt to physical properties through graph conditioning, we move closer to a regime where a single policy repository serves diverse robotic ecosystems.

The approach also informs broader cross-embodiment research beyond grasping. The principles of morphology-aligned graph representations and physical property injection could extend to legged locomotion (transferring across quadruped morphologies) or mobile manipulation (adapting to different arm configurations). The key insight, that structural alignment plus physical conditioning enables zero-shot transfer, may generalize to any domain where embodiments share topological regularities but vary in physical parameters.

Future work might explore continual learning scenarios where the policy expands its repertoire as new hand designs emerge, or investigate whether the graph representation can accommodate soft robotics through learned compliance parameters rather than rigid kinematic constraints. The release of code and datasets by the authors will undoubtedly accelerate progress in these directions.

Conclusion

DexGrasp-Zero demonstrates that zero-shot cross-embodiment dexterous grasping is achievable through careful attention to morphological representation and physical constraint integration. By replacing the error-prone retargeting pipeline with direct action mapping via morphology-aligned graphs, Wu et al. achieve substantial improvements in transfer performance. The 85% simulation success rate and 82% real-world performance on previously unseen hands suggest we are approaching practical universal grasping controllers.

The work raises important questions about the limits of structural alignment. How morphologically diverse can embodiments become while still sharing sufficient graph structure for transfer? Can we extend these principles to full-body humanoid control where morphology variations include limb counts and arrangements? As the robotics community continues to diversify hardware platforms, DexGrasp-Zero provides a compelling template for developing software that keeps pace with hardware innovation without requiring parallel retraining efforts for each new design.