irvinh/canvas-autonomous-explore-test

该数据集是通过[LeRobot](https://github.com/huggingface/lerobot)和[canvas-autonomous-learner](https://github.com/)协调器创建的，用于SO-101臂的小型自主学习循环。数据集包含2026-04-12一次EXPLORE爆发的2个episode样本，用于测试上传管道。数据集的具体描述包括：许可证为apache-2.0；使用的机器人是SO-101 follower（6个关节）；帧率为10 FPS；摄像头配置包括基础摄像头（640x480）和左手腕摄像头（640x480），均采用AV1编码；episodes为2个（一个正向肩部平移爆发，一个负向肩部平移爆发）；数据来源是通过`robotic-foundation-model-tests/scripts/run_single_action_record.py`脚本记录的自主EXPLORE爆发，该脚本包装了`lerobot.scripts.lerobot_record`并使用了自定义的`SingleActionPolicy`和每episode的任务描述。每个episode的自然语言任务（在`meta/tasks.parquet`中）格式为`"Move <joint> <direction> by <delta> units"`。此外，还有一个特定于仓库的辅助文件`meta/discrete_action_logs/episode_XXXXXX.jsonl`记录了每帧的离散动作命令以及头部信息，包括`joint_name`、`position_delta`、`action_duration`和`direction`。

This dataset was created using [LeRobot](https://github.com/huggingface/lerobot) via the [canvas-autonomous-learner](https://github.com/) orchestrator (a small autonomous-learning loop for the SO-101 arm). It is a 2-episode sample from one EXPLORE burst on 2026-04-12, pushed for testing the upload pipeline. The dataset description includes: License is apache-2.0; Robot is SO-101 follower (6 joints); FPS is 10; Cameras are base (640x480), left wrist (640x480), both AV1-encoded; Episodes are 2 (one positive-direction, one negative-direction shoulder_pan burst); Source is autonomous EXPLORE burst recorded via `robotic-foundation-model-tests/scripts/run_single_action_record.py`, which wraps `lerobot.scripts.lerobot_record` with a custom `SingleActionPolicy` and per-episode task descriptions. Each episodes natural-language task (in `meta/tasks.parquet`) is of the form `"Move <joint> <direction> by <delta> units"`. A repo-specific side-car `meta/discrete_action_logs/episode_XXXXXX.jsonl` also records the exact per-frame discrete action commanded by the policy, plus a header line with `joint_name`, `position_delta`, `action_duration`, and `direction`.