|
|
--- |
|
|
license: cc |
|
|
base_model: |
|
|
- Qwen/Qwen2.5-7B |
|
|
- Dexmal/Dexbotic-Base |
|
|
--- |
|
|
|
|
|
This model provides an initialization checkpoint for fine-tuning Dexbotic CogACT Model |
|
|
|
|
|
|
|
|
| Model | Description | Input Images | Action Dim | Model Size | |
|
|
| - | - | - |- | - | |
|
|
| Dexbotic-CogACT-HArm | Dual-arm CogACT model with multiple views input | Main View + Left Hand-View + Right Hand-View | 16D | 7B | |
|
|
|
|
|
|
|
|
## Action Dimension Description |
|
|
|
|
|
We recommend using the following data formats to fully utilize the pretrained models: |
|
|
|
|
|
1. **Single-arm end-effector pose**: Organize 7D action data as `[xyz + rpy + gripper]` |
|
|
2. **Single-arm joint angles**: Organize 8D action data as `[joints + gripper]` |
|
|
3. **Dual-arm end-effector pose**: Organize 14D action data as `[left_arm_xyz + left_arm_rpy + left_arm_gripper + right_arm_xyz + right_arm_rpy + right_arm_gripper]` |
|
|
4. **Dual-arm joint angles**: Organize 16D action data as `[left_arm_joints + left_arm_gripper + right_arm_joints + right_arm_gripper]` |
