| --- |
| language: |
| - en |
| - zh |
| tags: |
| - robotics |
| - teleoperation |
| - manipulation |
| - franka |
| - gello |
| size_categories: |
| - n<1K |
| pretty_name: Real-world robot teleoperation tasks |
| --- |
| |
| # Real-world robot teleoperation tasks / 真实机器人遥操作任务 |
|
|
| Raw GELLO teleoperation recordings from a Franka FR3 with a Robotiq 2F-85 gripper. Recordings include joint-space actions, measured robot state, and exterior/wrist camera videos at a nominal 15 Hz. |
|
|
| ## Tasks / 任务 |
|
|
| | Task / 任务 | Directory | Episodes | Samples | Duration | Raw size | |
| |---|---|---:|---:|---:|---:| |
| | 把番茄放进盒子 | [place_tomato_in_box](place_tomato_in_box) | 47 | 19,257 | 21.37 min | 162.56 MB | |
| | 把绿色方块叠到黄色方块上 | [stack_green_cube_on_yellow_cube](stack_green_cube_on_yellow_cube) | 117 | 45,040 | 49.98 min | 349.25 MB | |
| | 把黄色花朵插入花瓶 | [place_yellow_flower_in_vase](place_yellow_flower_in_vase) | 30 | 13,094 | 14.54 min | 118.01 MB | |
|
|
| Total: **194 episodes, 77,391 samples, 85.88 minutes, 629.82 MB** of original recordings (1,358 files). Collection dates: 2026-10-27 through 2026-11-02 (UTC). |
|
|
| The directories were renamed from `box_and_tomato`, `green_yellow_cube`, and `yellow_flower` to describe the intended task. The original episode IDs and all recording files are preserved byte-for-byte. Task names were inferred from sampled videos; no per-episode success labels were added. |
|
|
| ## Layout |
|
|
| ```text |
| <task_id>/ |
| README.md |
| task.json |
| episodes.csv |
| <episode_id>/ |
| metadata.json |
| samples.jsonl |
| trajectory.npz |
| videos/ |
| exterior.mp4 |
| exterior.frames.jsonl |
| wrist.mp4 |
| wrist.frames.jsonl |
| ``` |
|
|
| `episodes.csv` indexes recordings by ID, timestamp, sample count, duration, recording status, and video status. `manifest.json` lists file sizes, SHA-256 digests, and Git blob hashes for upload verification; the manifest itself is excluded from its file list. |
|
|
| ## Recording format |
|
|
| This repository preserves the project's raw recording format; it is not a converted LeRobot dataset. `trajectory.npz` can be opened with `numpy.load(path, allow_pickle=False)`. |
|
|
| | Array | Meaning | |
| |---|---| |
| | `action` (T × 8) | Seven absolute joint targets (radians), followed by gripper closedness | |
| | `proprioception` (T × 8) | Seven measured joint positions (radians), followed by measured gripper closedness | |
| | `joint_velocities` (T × 7) | Measured joint velocities | |
| | `eef_pose` (T × 7) | XYZ position in metres and WXYZ quaternion of the configured Franka end effector | |
| | `gello_input` (T × 8) | GELLO input before velocity limiting, including gripper closedness | |
| | `time_seconds` | Monotonic time relative to the episode start | |
| | `phase` | `recording` or `stopping` | |
| | `action_valid`, `proprioception_valid`, `gello_input_valid` | Masks identifying valid values; missing values may use zero placeholders | |
| | `command_submitted`, `command_id`, `submission_status` | Command submission tracking | |
|
|
| Gripper convention: **0 = open, 1 = closed**. Optional missing timestamps may be NaN. `samples.jsonl` also retains the submitted short trajectories and gripper events. `metadata.json` contains hardware configuration, recording statistics, and the final observation. |
|
|
| Video frames and numerical samples share a monotonic start time but are **not index-synchronized**. Align them using `time_seconds` and the frame JSONL records, which also contain source timestamps and duplicate/placeholder/gap flags. ROS source timestamps must not be directly subtracted from monotonic timestamps. Filter `phase == "recording"` if stopping-tail samples are unwanted. |
|
|
| All 194 recordings have capture status `completed`, while success labels are `null`. Capture completion does not assert manipulation success. Video quality flags from the original metadata and frame logs are retained; consult them when selecting or aligning data. |
|
|
