Datasets:
Modalities:
Video
Languages:
English
Size:
100K - 1M
ArXiv:
Tags:
robotics
reinforcement-learning
real-world-robotics
embodied-ai
teleoperation
human-intervention
Document SO101 dual-arm release
Browse files
README.md
CHANGED
|
@@ -29,7 +29,7 @@ RW-RL Dataset is a real-world robot interaction dataset released by Boden Intell
|
|
| 29 |
|
| 30 |
The broader RW-RL data program targets 1000+ hours of real robot interaction data across multiple robot embodiments, scene domains, task templates, and data regimes. In addition to teleoperated demonstrations, the dataset design emphasizes human intervention, autonomous rollouts, reward signals, done labels, and structured task semantics so that researchers can study policy correction, recovery, and closed-loop improvement in real environments.
|
| 31 |
|
| 32 |
-
> The current Hugging Face repository contains R1Lite and
|
| 33 |
|
| 34 |
## Dataset Coverage and Demo
|
| 35 |
|
|
@@ -168,6 +168,34 @@ The eight G2A camera streams are:
|
|
| 168 |
- `observation.images.camera_wrist_left`
|
| 169 |
- `observation.images.camera_wrist_right`
|
| 170 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 171 |
## File Layout
|
| 172 |
|
| 173 |
Each task folder is organized as follows:
|
|
|
|
| 29 |
|
| 30 |
The broader RW-RL data program targets 1000+ hours of real robot interaction data across multiple robot embodiments, scene domains, task templates, and data regimes. In addition to teleoperated demonstrations, the dataset design emphasizes human intervention, autonomous rollouts, reward signals, done labels, and structured task semantics so that researchers can study policy correction, recovery, and closed-loop improvement in real environments.
|
| 31 |
|
| 32 |
+
> The current Hugging Face repository contains R1Lite, G2A, PiperX, and SO101 releases. The full RW-RL program is designed as an expanding real-robot RL data base, and future subsets can follow the same file layout.
|
| 33 |
|
| 34 |
## Dataset Coverage and Demo
|
| 35 |
|
|
|
|
| 168 |
- `observation.images.camera_wrist_left`
|
| 169 |
- `observation.images.camera_wrist_right`
|
| 170 |
|
| 171 |
+
## SO101 Dual-Arm Release
|
| 172 |
+
|
| 173 |
+
The SO101 release contains six dual-arm manipulation tasks in LeRobot v3 format. It uses the
|
| 174 |
+
`bi_so_follower` robot type, 30 FPS H.264 video at 640 x 480, and 12-dimensional `action` and
|
| 175 |
+
`observation.state` vectors. Each task folder contains one parquet file per episode and one video
|
| 176 |
+
file per retained camera stream per episode.
|
| 177 |
+
|
| 178 |
+
| Task folder | Episodes | Video files | Camera streams | Hours |
|
| 179 |
+
| --- | ---: | ---: | ---: | ---: |
|
| 180 |
+
| `so101_insert_spray_pump_bottle_neck` | 251 | 753 | 3 | 15.131 |
|
| 181 |
+
| `so101_fold_clothes_left_stack_right` | 37 | 111 | 3 | 3.566 |
|
| 182 |
+
| `so101_return_tangram_pieces_original_state` | 145 | 435 | 3 | 4.832 |
|
| 183 |
+
| `so101_plug_power_cord_socket` | 44 | 88 | 2 | 2.790 |
|
| 184 |
+
| `so101_put_stationery_table_into_bag` | 32 | 96 | 3 | 0.929 |
|
| 185 |
+
| `so101_put_chopsticks_cups_storage_bucket` | 39 | 117 | 3 | 2.207 |
|
| 186 |
+
| **SO101 total** | **548** | **1,600** | - | **29.455** |
|
| 187 |
+
|
| 188 |
+
The three-camera tasks expose:
|
| 189 |
+
|
| 190 |
+
- `observation.images.left_wrist`
|
| 191 |
+
- `observation.images.right_front`
|
| 192 |
+
- `observation.images.right_wrist`
|
| 193 |
+
|
| 194 |
+
For `so101_plug_power_cord_socket`, one accepted source episode did not contain a historical
|
| 195 |
+
`right_wrist` stream. The release therefore uses the common two-camera intersection
|
| 196 |
+
(`left_wrist` and `right_front`) for all 44 episodes in that task. All episodes were retained, and
|
| 197 |
+
no synthetic, duplicated, or black replacement frames were introduced.
|
| 198 |
+
|
| 199 |
## File Layout
|
| 200 |
|
| 201 |
Each task folder is organized as follows:
|