| --- |
| language: |
| - en |
| tags: |
| - lerobot |
| - robotics |
| - internvla |
| - tactile |
| --- |
| # debug: robot demonstrations |
|
|
| **Instruction:** Nest the three paper cups together into a single stack. |
|
|
| LeRobot v3.0 dataset: **51 episodes, 43257 frames**, nominal 30 Hz. |
| Robot schema: `intern_gello_7dof_robotiq`. Original source recordings are retained by the owner. |
|
|
| ## Use |
|
|
| ```python |
| from lerobot.datasets.lerobot_dataset import LeRobotDataset |
| dataset = LeRobotDataset("cloudfan/intern-debug-lerobot", video_backend="torchcodec") |
| sample = dataset[0] |
| print(sample["task"], sample["observation.state"].shape, sample["action"].shape) |
| ``` |
|
|
| Copy `intern_gello_7dof_robotiq.yaml` into the InternVLA A-series training project's |
| `src/lerobot/dataset_schemas/configs/` directory. Use `action_mode=abs` and the dataset's |
| own normalization statistics initially. The schema maps two RGB views; tactile/force |
| channels are retained as auxiliary data, not automatically consumed by the RGB baseline. |
|
|
| ## Fields |
|
|
| - `observation.state`: 7 measured joint angles + measured Robotiq knuckle angle (radians). |
| - `action`: 7 absolute GELLO joint targets + commanded gripper open fraction (0 closed, 1 open). |
| Actions and observations use the same recorded frame, without an artificial temporal shift. |
| - `observation.images.cam_high`, `observation.images.cam_front`: 480 x 640 RGB. |
| - `observation.images.tactile_left_aug_diff`, `observation.images.tactile_right_aug_diff`: 700 x 400 RGB touch images. |
| - `observation.wrench`, `observation.wrench_valid`: force/torque and per-axis validity flags. |
| 0 episodes have missing force readings; only missing entries are filled with zero. |
| - `observation.eef_pose`: original six-value pose without inferred coordinate conversion. |
| - `observation.gripper_command`: original recorded follower gripper command. |
| - `source_timestamp`: original float64 timestamps retained in parquet. |
|
|
| Original JPEG packets are remuxed into MJPEG MP4 with no resize or re-encoding. |
| Every recorded frame is retained; standard timestamps are `frame_index / 30` rather than |
| wall-clock resampling. The actual capture FPS is available in the source manifest. |
| The gripper state (angle) and action (open fraction) have different physical meanings; |
| for delta control, apply subtraction only to the seven arm joints, using mask `[7, -1]`. |
|
|
| ## Quality and provenance |
|
|
| All source episodes are retained without success filtering. The prompt describes the intended |
| goal, not an annotation that every demonstration succeeded. Numeric fields were compared to |
| source values, JPEG bytes/order were SHA256-checked per stream, and first/middle/last frames |
| in every episode were decoded through InternVLA's LeRobot loader. See `meta/validation.json`, |
| `meta/source_manifest.json`, and `meta/conversion.json` for the full record. |
|
|
| Validated with [InternVLA A-series](https://github.com/InternRobotics/InternVLA-A-series/tree/e6fc904f9edbfb14532e97095fc2372202517f76). |
|
|
