Generalized Retargeting for Dexterous Manipulation

A Virtual Controller Framework Across Dataset and Teleoperation

IROS 2025 Hangzhou October 2025

Sponsored by

Abstract

We present a generic policy framework for anthropomorphic dexterous hand manipulation that bridges the gap between human demonstrations and robotic execution. Our framework learns robust, goal-directed manipulation behaviors that generalize across diverse tasks, while preserving human-like grasp and motion structure from limited demonstrations.

Key Contributions

  • Unified Retargeting Framework: Single framework applicable to both teleoperation and dataset-driven learning
  • Virtual Operator Pipeline: Introduces a learnable policy that simulates human teleoperation
  • Contact-Aware Retargeting: Addresses high-dimensional policy spaces with hardware constraints
  • Real Hardware Deployment: Successfully deployed on DexHand021 with JAKA Mini robot arm

Demonstration Videos

Power Grasp Regime

Whole-hand force closure demonstration

Enveloping Grasp

Full-wrap stabilization technique

Tool-Handle Grasp

Axis-aligned manipulation for tools

Note: Replace YouTube embed codes above with your actual video IDs, or use local video files:

<video controls width="100%">
    <source src="videos/demo1.mp4" type="video/mp4">
</video>

Experimental Results

Box Manipulation Results

Box Manipulation

ADD-AUC Score: 0.85

Mixer Manipulation Results

Mixer Operation

ADD-AUC Score: 0.82

Performance Metrics

We evaluated our model using the Area Under the Curve of Average Distance (ADD-AUC) metric across multiple YCB objects.

  • Average success rate: 78.5%
  • Real-time performance: 30 Hz control loop
  • Training datasets: HO3D, Arctic, DexYCB, DexCanvas

Downloads & Resources

📄

Workshop Poster

High-resolution PDF (3.4 MB)
💾

Supplementary Materials

Coming Soon
🔗

GitHub Repository

Source code and data

Citation

@inproceedings{iros2025workshop,
    title={Generalized Retargeting for Dexterous Manipulation},
    author={Anonymous Submission},
    booktitle={IROS 2025 Workshop on Dexterous Manipulation},
    year={2025},
    location={Hangzhou, China}
}

Contact

For questions about this work, please contact:

📧 your.email@example.com