Cross Embodied Co-Design for Dexterity

Co-Design Blind Policy

Baseline

We present Cross Embodied Co-Design – a novel approach for generating robot hands and dexterous control policies for real world deployment. Our method achieves dexterous manipulation by generating novel robot hand designs and evlauting with design conditioned, cross-embodied control - helping co-design for complex manipulation become more computationally tractable. We demonstrate dexterity by comparing four co-designed robot hands on unseen objects for in-hand rotation. Grasping, flipping, and in-hand rotation are tested in simulation. We achieve zero-shot sim-to-real transfer through physically grounded design grammars on a new, modular robot hand platform.

Paper

Latest Version (Dec. 30, 2025): here

Codebase (Coming Soon) Hardware Guide (Out Now!) Real World Demos

Key Features

🚀 Fast Training

A new robot hand can be generated, assembled, and deployed in under 24 hours.

🎯 Realistic Generation

Co-designed hands are generated with accurate collision meshes, wide range of morphology variations, and generalize to unseen objects.

🔧 Easy Sim-to-Real

Easy to build, assemble, and deploy with minimal debugging.

How It Works

Our approach consists of three main components: A scaleable cross-embodied policy, robot hand generation, and a modular robot hand platform. Each component helps create tractable sim-to-real for long horizon, dexterous tasks across morphologies. We leverage Graph Heuristic Search^[1], altered for cross embodied and dexterous hand evaluation, for search between ideal design and control pairs.

Generation: Modular grammars are used to construct varying families of robot hands - these include three groups of anthropomorphic and three groups of non-anthropomorphic hands. Hands are generated from primitive components with pre-computed collision meshes and dynamically computed convex hulls used for palms.
Cross Embodied Policy: Action masking is used to create a morphology conditioned cross embodied policies that can evaluate generated robot hands. This allows efficient evaluation across varying degrees of freedom, finger number, finger lengths, and fingertip types.
Robot Hand Platform: To deploy sim-to-real easily, a modular platform is created with modular electronics and servo motors.
Design Analysis: We provide initial analysis into design parameters of impact, showing morphology is critical to control.

Demonstrations

Resting Rotation

Rotate a ramdom object from a set of 16 YCB objects on the x axis.

Grasping

Grasp the object with a fixed wrist, hold until reset.

Flipping

Rotate an object resting on the ground about the z axis with a fixed wrist.

Generalization

We test our cross embodied policy and designs completely blind in the real world. All objects are unseen at test time with no state or tactile information given to the policy. We show the ability four our optimal co-designed hand to generalize across rigid and soft objects, object shape, weight, and textures. Our method is tested sim-to-real across four co-designed robot hands to show performance of the optimal found design relative to other co-designed robot hands and to show trade offs in design and control. We find the ranking and success in simulation can accurately predict real performance.

Hardware

Our hardware design is optimized for easy assembly with modular fingertips, degrees of freedom, finger number, palms, finger length, and finger placement. The system uses readily available components and can be assembled in under 4 hours without any specialty tools - only bolts, screw driver, and 3d printer required. Electronics are also modular using just one control board. The modular design allows for easy customization for additional co-design projects in manipulation.

💡

Pro Tip:

You can follow our hardware guide here for assembly tutorials, deployment code, and printing instructions.

Design Analysis

We explore physical parameters of impact to understand what parameters most affect control. Across all tested parameters, morphology parameters mattered most. Additional analysis of control limits and control type for long horizon tasks are provided in the paper.

Key Findings

Through our experiments on cross-embodied co-design for dexterous manipulation, we find the following to be of key importance:

Cross-embodiment enables tractable scaling: Cross-embodiment evaluation provides a practical approach to assess designs for complex manipulation without sacrificing fine control for computational speed, making co-design optimization feasible for real-world deployment.
Realistic grammar-based design enables zero-shot transfer: Grounding design grammars in physical components allows for seamless sim-to-real transfer without additional fine-tuning, as the modular structure maintains consistent physical properties across domains.
Morphology dominates the design space: Among all design parameters, morphological choices have the largest impact on manipulation performance, suggesting that shape and structure optimization should be prioritized over other design considerations.

Team

Kehlani Fay¹

Darin Djapri¹

Anya Zorin¹

James Clinton²

Ali El Lahib¹

Hao Su¹

Michael T. Tolley¹

Sha Yi¹

Xiaolong Wang¹

¹ University of California, San Diego ² University of California, Santa Barbra

Citation

@article{fay2025crossembodied, title={Cross Embodied Co-Design for Dexterous Hands}, author={Fay, Kehlani and Djapri, Darin and Zorin, Anya and Clinton, James and El Lahib, Ali and Su, Hao and Tolley, Michael T. and Yi, Sha and Wang, Xiaolong}, journal={arXiv preprint}, year={2025}, month={December} }

Cross-Embodied Co-Design for Dexterous Hands