AdityaShah/self_repair_gripper_screwdriver_bc

Flex-pi screwdriver pickup and tightening crops

Local LeRobot v2.1 dataset containing 800 episodes, 227,481 frames, 126.38 minutes at 30 Hz. Open review.html to browse every clip and switch between overhead, left-wrist, and right-wrist cameras.

Task: Pick up the screwdriver and tighten the screw securing the gripper in its holder.

Contents

• Original 32-dimensional observation.state and action, preserved bit-exact for retained rows. Dimension names and units follow the source; the feature names are in meta/info.json.
• All three RGB views, 640 × 360, synchronized to the retained rows. H.264 videos are re-encoded with CRF 18 to cut at arbitrary frames; RGB pixels are therefore not lossless copies.
• Frame indices and timestamps reset per episode, contiguous global indices, a new task instruction, camera intrinsics, and recomputed numerical and RGB statistics.
• meta/crop_manifest.json maps every output episode back to its source episode and half-open source frame interval [start_frame, end_frame_exclusive).
• meta/validation_report.json records the completed integrity audit; meta/loader_report.json records checks with the official LeRobot loader.

Depth videos are not included. Both arms' original state and action dimensions are retained; there is no action-space conversion or change to source observation/action timing.

Phase selection and review

Episode 0 is cropped from 31.53 to 38.90 seconds, frames [946, 1167). The crop includes the approach to grasp the screwdriver, pickup, transport, and tightening at the gripper holder, ending around the start of withdrawal. It excludes earlier gripper/screw handling and later tool return.

Each episode received its own kinematic proposal, using right gripper closure, low screwdriver pickup height, raised tool orientation, and holder approach. The start precedes grasp by up to 1.25 seconds; the end follows departure from the holder region by roughly 0.15 seconds. Four overhead frames (start, grasp, driving, end) were visually checked in every retained episode, with extra full-episode/detail sheets for exceptional trajectories. These are approximate phase boundaries, not dense frame-by-frame human annotations. Mechanical success and final screw torque are not labeled or guaranteed.

Source episodes 16 and 634 end before a complete screwdriver stage and are excluded. Source episode 147 has two tool attempts; only the final attempt is retained. Source episode 212 has overhead camera motion during the crop and is flagged in the manifest.

Load for behavior cloning

The dataset was checked with the official LeRobot v0.3.3 loader, whose dataset format is v2.1, datasets==3.6.0, and the PyAV video backend. Use a LeRobot release compatible with v2.1, or migrate with its conversion tools before using a loader that only accepts newer formats.

In a separate Python environment, install the compatible release with python -m pip install "lerobot==0.3.3" "datasets==3.6.0".

import json
from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = Path("data/flexpi_screwdriver/dataset").resolve()
splits = json.loads((root / "meta/recommended_splits.json").read_text())
train = LeRobotDataset(
    repo_id="local/flexpi-screwdriver",
    root=root,
    episodes=splits["train"],
    video_backend="pyav",
    delta_timestamps={"action": [i / 30 for i in range(32)]},
)
sample = train[0]
# RGB: (3, 360, 640); state: (32,); action chunk: (32, 32).
# Respect sample["action_is_pad"] in the loss at episode ends.

meta/recommended_splits.json contains 725 training and 75 validation episodes. Exact duplicates of source state/action trajectories stay together: source pairs 155/303, 156/304, 157/305, 158/306, 159/307, and 160/308. All six extra copies are retained. Pass the split episode lists explicitly; the standard metadata train split contains all 800 episodes. This prevents leakage from the detected exact duplicates; it is not a recording-session or operator-level holdout.

Provenance and license

Derived from Flex-pi's self_repair_gripper_bc dataset, published by Flex-pi, pinned at revision 3f5903e8162ddd1e42035d279b6f687ae191afb6. The source is licensed CC BY 4.0; this cropped derivative retains that license and attribution. Changes comprise temporal selection, RGB re-encoding, episode/time/index remapping, a stage-specific task label, regenerated statistics, annotations, and documentation. No source state or action values were modified within retained rows.

Rebuilding and validation scripts are in tools/flexpi_screwdriver/ in the project. The adjacent source/ directory contains the pinned downloaded data; the dataset/ folder itself is portable and contains everything needed to load these crops.

Load from Hugging Face

The public dataset is here. For a small first-episode check, the compatible LeRobot loader downloads just the required episode tables and RGB videos:

from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset(
    "AdityaShah/self_repair_gripper_screwdriver_bc",
    episodes=[0],
    video_backend="pyav",
    delta_timestamps={"action": [i / 30 for i in range(32)]},
)
sample = dataset[0]

For training, download meta/recommended_splits.json using hf_hub_download(repo_id="AdityaShah/self_repair_gripper_screwdriver_bc", filename="meta/recommended_splits.json", repo_type="dataset") and pass its training or validation episode list to episodes. The repository has a v2.1 format tag for the compatible loader.