GR-Dexter: VLA for Bimanual Dexterous Robot Control
Organization: ByteDance Research
Overview
GR-Dexter is an integrated framework for enabling Vision-Language-Action (VLA) models to control bimanual dexterous robots. The system addresses challenges in scaling robotic manipulation beyond simple grippers to high-degree-of-freedom hands.
Key Hardware Components
ByteDexter V2 Hand:
- 21 degrees of freedom (four per finger, five for thumb)
- Dimensions: 219mm height × 108mm width
- High-density piezoresistive sensor arrays for tactile feedback across fingertips
Teleoperation Interface (data collection):
- Meta Quest VR for wrist tracking
- Manus Metagloves for hand capture
- Foot pedals for arm control
- Enables coordination of two Franka arms for long-horizon manipulation tasks
Model Architecture
- Parameters: 4 billion
- Architecture: Mixture-of-Transformer (MoT)
- Inputs: language instructions, visual observations, robot state
- Outputs: action chunks controlling arm joints, end-effector poses, hand joints, fingertip positions
Training Strategy
Co-training on three data sources:
- Web-scale vision-language datasets
- Cross-embodiment robot data (Fourier ActionNet, OpenLoong, RoboMIND)
- Human VR trajectories (with temporal consistency filtering)
Performance Results
| Task | Condition | Success Rate |
|---|---|---|
| Long-Horizon Tasks | In-distribution | 0.97 |
| Long-Horizon Tasks | Out-of-distribution | 0.89 |
| Long-Horizon Tasks (baseline) | Out-of-distribution | 0.64 |
| Pick-and-Place | Unseen objects | 0.85 |
| Pick-and-Place | Unseen instructions | 0.83 |
Cross-embodiment data significantly improves generalization to novel scenarios while maintaining strong in-domain performance.