physical-ai-toolchain/ur10e-gear-pick-place-demos

UR10e Gear Pick-and-Place Demonstrations

Private archive of the exact LeRobot v3.0 train and validation directories used by the revised Rho delta-action IL run. The task is to pick up the large gear and place it in the bin using a simulated UR10e with a Robotiq 2F-140 gripper.

📊 Dataset identity

The corpus was collected on September 2, 2026 under privileged-gear-dual-camera-100-reduced-release-20260902. These are accepted privileged-expert demonstrations, not learned-policy rollouts or human-recorded real-world demonstrations. One source episode per ten-episode collection batch was reserved for validation.

📦 Layout and loading

Each of train/ and validation/ is a complete, independently loadable LeRobot root with its own meta/, data/, and videos/. Download the repository snapshot and pass the selected subdirectory as the local dataset root. The repository root itself is not a single LeRobot dataset root.

Each original subdirectory retains its own internal meta/info.json split named train; the outer directory names distinguish experiment training from validation. The Hugging Face card configuration exposes the two sets of Parquet records as train and validation without rewriting the original LeRobot metadata. Generic Parquet loading does not replace LeRobot's synchronized video decoding.

Original provenance/configuration files retain their workstation paths and local dataset IDs. Treat those as historical identifiers, not portable paths. Override roots with the downloaded train/ and validation/ directories, use config/train-stats-chunk50.json for both splits, and resolve priority manifests from config/. Do not regenerate splits or statistics merely to load the archived data.

🎥 Observation and action contract

The dataset stores absolute targets. Rho training transforms all seven targets, including the gripper, into deltas relative to the current raw measured state, then applies per-horizon action-chunk quantile normalization. Images are mapped to two 256 × 256 policy inputs according to the saved training recipe.

The physical simulation uses 120 Hz physics and 30 Hz control, source object offsets within ±15 mm per XY axis, and a slow partial gripper opening for in-bin release. The two cameras contain the original occlusions and dark wrist-camera frames.

⚠️ Interpretation and usage

The 100 accepted episodes are a finalized filtered corpus, not a 100% success estimate for all collection attempts. Validation episodes were excluded from the revised IL training run. Do not split individual frames randomly across training and validation.

Grasp-transition priority sampling selected 14,940 training indices and targeted a 25% sampling fraction. It emphasized close, contact dwell, micro-lift, and early lift, not release windows.

[!WARNING] The associated historical checkpoint was trained with a scalar STATE-normalization defect. The raw dataset is preserved unchanged; the model repository supplies the recorded input profile for reproducing that checkpoint. New corrected-normalization training must be a separately identified experiment.

Raw HDF5 evidence, the duplicate unsplit dataset, earlier policy checkpoints, and PPO trajectories are excluded. No blanket third-party asset license is asserted by this private archive. The data provides no evidence of real-world deployment safety or sim-to-real success.

📚 Related artifacts

• —Simulator: physical-ai-toolchain/ur10e-gear-pick-place-sim (dataset).
• —Trained policy: physical-ai-toolchain/rho-ur10e-gear-pick-place-il-delta-25500 (model).