--- license: apache-2.0 task_categories: - robotics tags: - LeRobot - so101 - so-arm101 - robotics - manipulation - grasping - act - real-robot configs: - config_name: default data_files: data/*/*.parquet --- This dataset was created using [LeRobot](https://github.com/huggingface/lerobot). ## Dataset Description 118 teleoperated demonstrations of an [SO-101](https://github.com/TheRobotStudio/SO-ARM100) arm picking up a small tube from a mat and dropping it into a bin, recorded with two cameras (wrist and side) at 30 fps. **An ACT policy trained on this data reaches 96.7% success (29/30) on real hardware.** That number is the point of this dataset. Plenty of demonstration sets exist; few state what success rate the data actually produces, under what protocol, or how the data was assembled to get there. ## Measured result | Policy | Training episodes | Success rate | Protocol | |--------|------------------|--------------|----------| | v2 | 50 | 40% (8/20, reproduced twice) | 20 trials, 40 s | | v3 | 75 | 60% (24/40 pooled) | 2x20 trials, 40 s | | v4 | 107 | 82.5% (33/40 pooled) | 2x20 trials, 40 s | | **v5 (this dataset)** | **118** | **96.7% (29/30)** | 30 trials, 50 s | Trained with ACT for 57k steps. Success criterion: tube securely grasped and lifted, with tube positions deliberately varied across the workspace. ## How it was built Not one bulk recording session. The 118 episodes are 50 base demonstrations plus five rounds of *targeted corrections*, each round recorded after evaluating the current policy and identifying its specific failure mode. Architecture, hyperparameters and hardware were unchanged throughout -- only the data changed. The composition is the interesting part: episodes were added to cover workspace regions and grasp situations where the policy was measurably failing, rather than to increase the count. ## Usage ```python from lerobot.datasets.lerobot_dataset import LeRobotDataset ds = LeRobotDataset("Atomictan/grab_tube_merged_v3") ``` Train an ACT policy on it: ```bash lerobot-train \ --dataset.repo_id=Atomictan/grab_tube_merged_v3 \ --policy.type=act \ --output_dir=outputs/train/grab_tube \ --job_name=grab_tube \ --policy.device=cuda ``` It also works as an offline buffer for RL, though note it carries no reward column -- `ReplayBuffer.from_lerobot_dataset` requires `next.reward`, so you would need to label it first (real demonstrations have no ground-truth success signal). ## Known limitations - **Single task, single environment.** One tube, one mat, one lighting condition. Expect a policy trained on this to be sensitive to background changes; ACT fine-tunes its ResNet18 backbone on relatively few episodes and picks up incidental scene cues. - **Tube positions are concentrated** where the correction cycles focused. Coverage is good across the reachable workspace but not uniform. - **No reward labels.** Success/failure is not annotated per episode. - **`robot_type` is recorded as `so_follower`**, from a local fork's naming, rather than the more standard `so101_follower`. The data itself is stock SO-101. ## Provenance and method Recorded July 8-22, 2026 on a single hobbyist build, as part of learning robot ML end-to-end. The method worth copying is the measurement, not the data. Casual testing of the first policy suggested roughly 70% success; a fixed protocol -- 20 trials, deliberately varied tube positions, a strict success criterion -- measured 40%. Informal testing samples the easy positions the demonstrations already covered. From there each cycle was: evaluate under the fixed protocol, tag every failure with a category and workspace zone, form a hypothesis about the dominant failure mode, and record corrections targeting only that. One hypothesis (background/scene drift between recording and evaluation) was tested by restoring the original scene and re-running the protocol; it moved success 35% -> 40%, within noise, and was rejected before any training time was spent on it. The position-coverage hypothesis survived, and drove the remaining cycles. Full writeup, including the failure taxonomy and what each correction round targeted: [`GRASP_IMPROVEMENT_SUMMARY.md`](https://github.com/atomictan/lerobot-so101/blob/main/scripts/learning/GRASP_IMPROVEMENT_SUMMARY.md). The wider project log -- imitation learning, simulation, and reinforcement learning on the same arm -- is in [`PROJECT_LOG.md`](https://github.com/atomictan/lerobot-so101/blob/main/scripts/learning/PROJECT_LOG.md). - **Homepage:** https://github.com/atomictan/lerobot-so101 - **Paper:** [More Information Needed] - **License:** apache-2.0 ## Dataset Structure [meta/info.json](meta/info.json): ```json { "codebase_version": "v3.0", "fps": 30, "features": { "action": { "dtype": "float32", "names": [ "shoulder_pan.pos", "shoulder_lift.pos", "elbow_flex.pos", "wrist_flex.pos", "wrist_roll.pos", "gripper.pos" ], "shape": [ 6 ] }, "observation.state": { "dtype": "float32", "names": [ "shoulder_pan.pos", "shoulder_lift.pos", "elbow_flex.pos", "wrist_flex.pos", "wrist_roll.pos", "gripper.pos" ], "shape": [ 6 ] }, "observation.images.wrist": { "dtype": "video", "shape": [ 640, 480, 3 ], "names": [ "height", "width", "channels" ], "info": { "is_depth_map": false, "video.height": 640, "video.width": 480, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {} } }, "observation.images.side": { "dtype": "video", "shape": [ 480, 640, 3 ], "names": [ "height", "width", "channels" ], "info": { "is_depth_map": false, "video.height": 480, "video.width": 640, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.fps": 30, "video.channels": 3, "has_audio": false, "video.g": 2, "video.crf": 30, "video.preset": 12, "video.fast_decode": 0, "video.video_backend": "pyav", "video.extra_options": {} } }, "timestamp": { "dtype": "float32", "shape": [ 1 ], "names": null }, "frame_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "episode_index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "index": { "dtype": "int64", "shape": [ 1 ], "names": null }, "task_index": { "dtype": "int64", "shape": [ 1 ], "names": null } }, "total_episodes": 118, "total_frames": 95305, "total_tasks": 1, "chunks_size": 1000, "data_files_size_in_mb": 100, "video_files_size_in_mb": 200, "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet", "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4", "robot_type": "so_follower", "splits": { "train": "0:118" } } ``` ## Citation **BibTeX:** ```bibtex @misc{grab_tube_so101_2026, title = {Grab the Tube: 118 SO-101 demonstrations reaching 96.7% with ACT}, author = {Atomictan}, year = {2026}, url = {https://huggingface.co/datasets/Atomictan/grab_tube_merged_v3} } ```