simpleicl-sample

# SimpleICL Real-Robot Data Sample A sample of our paired human-robot manipulation data, released alongside [SimpleICL](https://github.com/simpleicl/simple-icl) — the official code release of the paper *In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks*. It is a subset of the full dataset collected for the project, provided to illustrate the data format and to accompany the training recipe in the repository README (*Training*). ## Dataset Summary 100 paired episodes of bimanual Franka FR3 tabletop manipulation. In every pair, a teleoperated robot arm and a human hand perform the same task back to back; the human-side demonstration video serves as the in-context condition (`video_ctx`) that the policy consumes at training and inference time. | | | | --- | --- | | Episodes | 100 pairs (100 robot + 100 human) | | Total frames (robot side) | 28,457 @ 30 fps | | Episode length | 157–449 frames | | Robot side | 3 RealSense cameras (high + left/right wrist), 848×480 | | Human side | 1 high camera, 848×480 | | Action / state | 16-D (7 joints + 1 gripper per arm) | Scene alterations keep the task unchanged and perturb only task-irrelevant factors — object positions, lighting, and background colors. The same task recorded under the base and altered placements references each other as in-context conditions (see `robot/meta/conditions_list.json`), so a policy trained on this data learns *what the task is* from the demonstration video while staying robust to where objects sit, how the scene is lit, and what the background looks like. ## Structure The data is distributed as a single archive, `pair_data.zip`. Unzipping it produces two sibling LeRobot-format (v2.1) dataset directories: ``` pair_data.zip └── (extracts to:) ├── robot/ # teleoperated robot episodes │ ├── meta/ # tasks, episodes, scenes, conditions │ │ ├── tasks.jsonl │ │ ├── episodes.jsonl │ │ ├── episodes_stats.jsonl │ │ ├── info.json │ │ ├── scenes.json # episode index -> scene placement label │ │ ├── plan_rows.json │ │ ├── conditions.json │ │ └── conditions_list.json # episode index -> human demo videos (video_ctx) │ ├── data/ # parquet per episode │ └── videos/ # 3 cameras per episode └── human/ # human-hand demonstrations └── videos/ # cam_high video per episode ``` Condition paths in `robot/meta/conditions_list.json` are relative to the extraction root, so the dataset keeps working after the extracted directory is moved as a whole. ## Usage Download the archive, unzip it, and point the SimpleICL training config at the robot side — see the repository [SimpleICL](https://github.com/simpleicl/simple-icl) README for the one-command training flow. ## Additional Information - **License**: MIT - **Origin**: a sample subset of the data collected by the SimpleICL authors on a dual-arm Franka FR3 platform with RealSense cameras. - **Associated code**: [SimpleICL](https://github.com/simpleicl/simple-icl)