PDF Paper ArXiv arXiv Youtube Video X (Twitter) Summary
Related Research
Power to Precision Dex1B

From Power to Precision

Learning Fine-grained Dexterity
for Multi-fingered Robotic Hands

Jianglong Ye*, Lai Wei*, Guangqi Jiang, Changwei Jing,
Xueyan Zou, Xiaolong Wang

UC San Diego Logo UC San Diego Logo
PDF Paper ArXiv arXiv Youtube Video X (Twitter) Summary
Power to Precision Dex1B

Recent studies have shown that dexterous hands work well for power grasp, but for precision oriented tasks, parallel grippers are still more common.🤔


We present Power💪 to Precision🎯, a co-design framework that optimizes both control and finger-tip geometry for multi fingered robotic hands to achieve both power and precision.

Recent studies have shown that dexterous hands work well for power grasp, but for precision oriented tasks, parallel grippers are still more common.🤔


We present Power💪 to Precision🎯, a co-design framework that optimizes both control and finger-tip geometry for multi fingered robotic hands to achieve both power and precision.

Technical Summary Video

Real-to-Real Deployment

Our framework enables teleoperation of fine grained manipulation tasks using dexterous hands and improves deployment robustness. The following video shows real to real deployment of the framework. Policies are trained with teleoperation data.

G1 Inspire

XArm XHand

Sim-to-Real Deployment

Along with the real-to-real deployment, we also showcase the sim-to-real deployment of the framework. Policies are trained with simulation data and exhibit better spatial & object generalizability.

XArm XHand

G1 Inspire

Comparison

Teleoperation

The advantage of our co design framework for teleoperation is clear when we compare our method with the retargeting baseline. All videos are shown at x4 speed.
DexRetargeting
Ours

Sim-to-Real Policy

Compared with Dex1B, our co design framework can precisely grasp small objects. All videos are shown at x4 speed.
Dex1B
Ours

Co-Design Framework

We introduce a co-design framework that jointly optimizes control and fingertip geometry for multi-fingered robotic hands, enabling both power and precise manipulation.

Control Optimization

Simplified contact with parallel finger motions greatly improves the robustness of the resulting actions.
Method (Control)

Fingertip Geometry Optimization

The fingertip geometry is modeled as a contact plane and optimized using a neural physics surrogate model.
Method (Design)

BibTeX

@article{ye2025powertoprecision,
  title={From Power to Precision: Learning Fine-grained Dexterity for Multi-fingered Robotic Hands},
  author={Ye, Jianglong and Wei, Lai and Jiang, Guangqi and Jing, Changwei and Zou, Xueyan and Wang, Xiaolong},
  journal={arXiv},
  year={2025},
}
Click to copy