Update README.md
Browse files
README.md
CHANGED
|
@@ -14,13 +14,99 @@ size_categories:
|
|
| 14 |
|
| 15 |
This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
|
| 16 |
|
| 17 |
-
#
|
| 18 |
|
|
|
|
|
|
|
| 19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
- **
|
| 22 |
-
- **
|
| 23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
## Dataset Structure
|
| 26 |
|
|
@@ -289,4 +375,6 @@ This dataset was created using [LeRobot](https://github.com/huggingface/lerobot)
|
|
| 289 |
"data_files_size_in_mb": 100,
|
| 290 |
"video_files_size_in_mb": 500
|
| 291 |
}
|
| 292 |
-
```
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
|
| 16 |
|
| 17 |
+
# 1 Dataset Description
|
| 18 |
|
| 19 |
+
This dataset includes 32 Lightwheel-Robocasa-Tasks, collected using the g1-wbc robot in environments provided by LW-BenchHub.
|
| 20 |
+
The robot configuration used during data collection is **`G1-Controller-DecoupledWBC`**, using `UnitreeG1ControllerDecoupledWBCEnvCfg` inherits from `UnitreeG1ControllerEnvCfg`.
|
| 21 |
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## 1.1 Robot State
|
| 25 |
+
|
| 26 |
+
The robot state recorded in the environment is stored in: **observation.state**
|
| 27 |
+
- **Dimension**: 43
|
| 28 |
+
- **Description**: Joint positions of the robot
|
| 29 |
+
- **Joint mapping**: The exact joint ordering and semantics can be found in the **Dataset Structure** section below.
|
| 30 |
+
|
| 31 |
+
---
|
| 32 |
+
|
| 33 |
+
## 1.2 Robot Action
|
| 34 |
+
|
| 35 |
+
The robot actions are recorded in: **action**
|
| 36 |
+
- **Dimension**: 23
|
| 37 |
+
The action is ordered as: left hand (1), right hand (1), left arm (7), right arm (7), base (7).
|
| 38 |
+
|
| 39 |
+
| Component | Dimension | Description |
|
| 40 |
+
|:----------------|:---------:|:------------------------------------|
|
| 41 |
+
| Hand action | 2 | Continuous gripper control |
|
| 42 |
+
| Arm action | 14 | Left arm + right arm (end-effector poses) |
|
| 43 |
+
| Base action | 7 | Base motion and torso orientation commands |
|
| 44 |
+
|
| 45 |
+
The g1-wbc robot uses `G1DecoupledWBCActionsCfg` to implement whole-body control for all components except the grippers.
|
| 46 |
+
```python
|
| 47 |
+
self.action_config = G1DecoupledWBCActionsCfg()
|
| 48 |
+
self.action_config.left_hand_action = self.gripper_cfg.left_hand_action_cfg()[self.hand_action_mode] # here self.hand_action_mode = "handle"
|
| 49 |
+
self.action_config.right_hand_action = self.gripper_cfg.right_hand_action_cfg()[self.hand_action_mode]
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
## 1.2.1 Gripper Action
|
| 53 |
+
|
| 54 |
+
Grippers are controlled using continuous values. In the action data, `-1.0` indicates fully open, `+1.0` indicates fully closed.
|
| 55 |
+
|
| 56 |
+
- **Action values**:
|
| 57 |
+
- `-1.0`: fully open gripper
|
| 58 |
+
- `+1.0`: fully closed gripper
|
| 59 |
+
- Continuous values in between represent partial opening/closing
|
| 60 |
|
| 61 |
+
- **Left gripper** (index 0): Continuous value from -1.0 (open) to 1.0 (closed)
|
| 62 |
+
- **Right gripper** (index 1): Continuous value from -1.0 (open) to 1.0 (closed)
|
| 63 |
+
|
| 64 |
+
- **Source**: VR trigger mapping
|
| 65 |
+
- `trigger=0.0` (not pressed) → `-1.0` (open)
|
| 66 |
+
- `trigger=1.0` (fully pressed) → `1.0` (closed)
|
| 67 |
+
|
| 68 |
+
## 1.2.2 Arm Action
|
| 69 |
+
|
| 70 |
+
The arm actions control the end-effector poses (position + orientation) in the robot frame.
|
| 71 |
+
The left and right arm actions are processed in `G1DecoupledWBCAction.process_actions()` and converted into 14-dimensional arm joint positions via inverse kinematics.
|
| 72 |
+
|
| 73 |
+
- **Left arm action** (indices 2-8):
|
| 74 |
+
- `left_arm_pos_x` (index 2): X position in robot frame
|
| 75 |
+
- `left_arm_pos_y` (index 3): Y position in robot frame
|
| 76 |
+
- `left_arm_pos_z` (index 4): Z position in robot frame
|
| 77 |
+
- `left_arm_quat_w` (index 5): Quaternion w component
|
| 78 |
+
- `left_arm_quat_x` (index 6): Quaternion x component
|
| 79 |
+
- `left_arm_quat_y` (index 7): Quaternion y component
|
| 80 |
+
- `left_arm_quat_z` (index 8): Quaternion z component
|
| 81 |
+
|
| 82 |
+
- **Right arm action** (indices 9-15):
|
| 83 |
+
- `right_arm_pos_x` (index 9): X position in robot frame
|
| 84 |
+
- `right_arm_pos_y` (index 10): Y position in robo frame
|
| 85 |
+
- `right_arm_pos_z` (index 11): Z position in robot frame
|
| 86 |
+
- `right_arm_quat_w` (index 12): Quaternion w component
|
| 87 |
+
- `right_arm_quat_x` (index 13): Quaternion x component
|
| 88 |
+
- `right_arm_quat_y` (index 14): Quaternion y component
|
| 89 |
+
- `right_arm_quat_z` (index 15): Quaternion z component
|
| 90 |
+
|
| 91 |
+
- **Coordinate frame**: Robot frame
|
| 92 |
+
|
| 93 |
+
## 1.2.3 Base Action
|
| 94 |
+
|
| 95 |
+
The base action controls base motion and torso orientation using task space commands in the local robot frame.
|
| 96 |
+
The base action commands are processed by `G1DecoupledWBCAction.process_actions()` and converted into 15-dimensional lower-body joint positions (15).
|
| 97 |
+
|
| 98 |
+
- **Controlled dimensions** (indices 16-22):
|
| 99 |
+
- `base_lin_x_local` (index 16): Forward/backward command in local frame (range: -1.0 to 1.0)
|
| 100 |
+
- `base_lin_y_local` (index 17): Left/right command in local frame (range: -1.0 to 1.0)
|
| 101 |
+
- `base_ang_z_local` (index 18): Rotation (yaw) command in local frame (range: -1.0 to 1.0)
|
| 102 |
+
- `base_height_cmd` (index 19): Base height command (typically 0.2 to 0.75, default 0.75 for standing)
|
| 103 |
+
- `torso_orientation_roll_cmd` (index 20): Torso roll orientation command (range: -0.5 to 0.5 (clamped), typically 0 in data)
|
| 104 |
+
- `torso_orientation_pitch_cmd` (index 21): Torso pitch orientation command (range: -1.5 to 1.5 (clamped), typically 0 in data)
|
| 105 |
+
- `torso_orientation_yaw_cmd` (index 22): Torso yaw orientation command (range: -2.0 to 2.0 (clamped), typically 0 in data)
|
| 106 |
+
|
| 107 |
+
- **Coordinate frame**: Robot frame
|
| 108 |
+
|
| 109 |
+
---
|
| 110 |
|
| 111 |
## Dataset Structure
|
| 112 |
|
|
|
|
| 375 |
"data_files_size_in_mb": 100,
|
| 376 |
"video_files_size_in_mb": 500
|
| 377 |
}
|
| 378 |
+
```
|
| 379 |
+
## Homepage: [[Lightwheel-BenchHub](https://docs.lightwheel.net/lw_benchhub/)]
|
| 380 |
+
## License: [[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)]
|