RoCo Task Board Assembly overview

RoCo Task Board Assembly Demonstrations

Real-world LeRobot demonstrations for contact-rich task-board assembly.

## Overview **RoCo Task Board Assembly Demonstrations** is a real-world robot manipulation dataset for assembling parts on a task board. It is released by **Sharpa** in a **LeRobot-compatible** format for imitation learning, visuomotor policy learning, visual-tactile representation learning, and contact-rich manipulation research. Task-board assembly requires precise part localization, fine contact timing, bimanual coordination, and robust perception under hand-object occlusion. The demonstrations include synchronized multi-view video, tactile observations, proprioceptive state, torque signals, and action targets.

Task Task-board part assembly	Format LeRobot v3.0 / v2.1	Scale 10 seasons / 151 episodes	Frequency 30 FPS
Video 6 synchronized streams	State / Action 65D joint space	Tactile 60D signal + tactile video	Use Training and policy development

## Dataset Capabilities | Capability | Dataset Support | | --- | --- | | Contact-rich assembly learning | Real-world demonstrations for assembling task-board parts | | Multi-view visuomotor policies | Synchronized head-camera and wrist-camera observations | | Visual-tactile learning | Tactile deformation video, raw tactile video, and 60D tactile signals | | Joint-space control | 65D synchronized state and action for two arms, two dexterous hands, and torso/motor-related joints | | LeRobot ecosystem | `lerobot3.0` and `lerobotv2.1` exports for every released season | ## Example Views The demonstrations include synchronized head, wrist, and tactile video streams. Each preview below uses a 3-minute window from the same `lerobotv2.1` episode, played at 10x speed. GIF previews render directly in Markdown; click any preview to open the MP4 version. Tactile previews preserve their full wide-frame layout.

Head Left	Head Right
Wrist Left	Wrist Right
Tactile Deformation	Raw Tactile

## Dataset Statistics | Item | Value | | --- | ---: | | Total collection seasons | 10 | | `lerobot3.0` seasons | 10 | | `lerobot3.0` episodes | 151 | | `lerobot3.0` frames | 733,809 | | `lerobotv2.1` seasons | 10 | | `lerobotv2.1` episodes | 151 | | `lerobotv2.1` frames | 733,809 | | FPS | 30 | | Video streams | 6 | | State/action dimension | 65 | | Tactile signal dimension | 60 | | Approximate data size | 104.8 GB | For new users, we recommend starting from `lerobot3.0`. The `lerobotv2.1` export is included for compatibility with pipelines that still depend on the older LeRobot layout. ## Get Started ### Download The Dataset Make sure Git LFS is installed before cloning from Hugging Face. ```bash git lfs install git clone https://huggingface.co/datasets/SharpaIT/RoCo_TaskBoardAssembly ``` If you want to clone only metadata first and fetch large files later: ```bash GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/SharpaIT/RoCo_TaskBoardAssembly ``` If you only want a specific season, use sparse checkout: ```bash git init RoCo_TaskBoardAssembly cd RoCo_TaskBoardAssembly git remote add origin https://huggingface.co/datasets/SharpaIT/RoCo_TaskBoardAssembly git sparse-checkout init git sparse-checkout set season_POC22061_2026_06_11_14_29_08_train README.md git pull origin main ``` ### Quick Inspection Each season contains both `lerobot3.0` and `lerobotv2.1` exports. Inspect `meta/info.json` first to understand the exact schema and file templates. ```python import json from pathlib import Path dataset_root = Path("RoCo_TaskBoardAssembly") episode_root = dataset_root / "season_POC22061_2026_06_11_14_29_08_train" / "lerobot3.0" with open(episode_root / "meta" / "info.json", "r") as f: info = json.load(f) print(info["total_episodes"]) print(info["total_frames"]) print(info["features"].keys()) ``` ## Dataset Structure The dataset is organized by collection season. Each season contains a `lerobot3.0` export and a `lerobotv2.1` export. ```text RoCo_TaskBoardAssembly/ ├── README.md ├── season_POC22061_2026_06_11_14_29_08_train/ │ ├── lerobot3.0/ │ │ ├── meta/ │ │ │ ├── info.json │ │ │ ├── modality.json │ │ │ ├── episodes/ │ │ │ └── tasks.parquet │ │ ├── data/ │ │ │ └── chunk-000/ │ │ └── videos/ │ │ ├── observation.images.head_left/ │ │ ├── observation.images.head_right/ │ │ ├── observation.images.wrist_left/ │ │ ├── observation.images.wrist_right/ │ │ ├── observation.images.tactile_deform/ │ │ └── observation.images.tactile_raw/ │ └── lerobotv2.1/ │ ├── meta/ │ ├── data/ │ └── videos/ └── season_.../ ``` ### LeRobot Storage Layout | Part | Description | | --- | --- | | `meta/` | Dataset metadata, feature schema, task metadata, and path templates | | `data/` | Episode frame data stored as Apache Parquet files | | `videos/` | Per-camera MP4 videos | The most important metadata file is `meta/info.json`. It defines `total_episodes`, `total_frames`, `fps`, `splits`, `data_path`, `video_path`, and `features`. ### File Path Templates LeRobot v3.0 uses templates similar to: ```text data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 ``` LeRobot v2.1 uses templates similar to: ```text data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4 ``` ## Features Schema ### Main Feature Groups | Feature | Type | Shape | Description | | --- | --- | ---: | --- | | `observation.state` | float32 | 65 | Joint-space robot state | | `action` | float32 | 65 | Joint-space action | | `observation.state.joint_torque` | float32 | 65 | Joint torque signal | | `observation.tactile` | float32 | 60 | Tactile force/torque signal | | `observation.images.*` | video | varies | Multi-view visual observations | | `timestamp` | float32 | 1 | Timestamp | | `frame_index` | int64 | 1 | Frame index within an episode | | `episode_index` | int64 | 1 | Episode index | | `task_index` | int64 | 1 | Task index | ### Proprioceptive State The 65D `observation.state` and `action` vectors are ordered as follows: | Range | Names | Meaning | | --- | --- | --- | | 0 - 6 | `left_arm_j0` to `left_arm_j6` | Left arm joints | | 7 - 28 | `left_hand_j0` to `left_hand_j21` | Left dexterous hand joints | | 29 - 35 | `right_arm_j0` to `right_arm_j6` | Right arm joints | | 36 - 57 | `right_hand_j0` to `right_hand_j21` | Right dexterous hand joints | | 58 - 64 | `motor_j0` to `motor_j6` | Torso / motor-related joints | ### Video Streams The complete visual observation set contains six camera streams: | Feature key | Description | Shape | | --- | --- | --- | | `observation.images.head_left` | Left head camera | 480 x 480 x 3 | | `observation.images.head_right` | Right head camera | 480 x 480 x 3 | | `observation.images.wrist_left` | Left wrist camera | 480 x 480 x 3 | | `observation.images.wrist_right` | Right wrist camera | 480 x 480 x 3 | | `observation.images.tactile_deform` | Tactile deformation video | 480 x 1200 x 3 | | `observation.images.tactile_raw` | Raw tactile video | 480 x 1600 x 3 | Video files are MP4 without audio. Codec may differ across exports and seasons. `lerobot3.0` is primarily AV1, while `lerobotv2.1` is primarily H.264. ## Season List | Season | `lerobot3.0` | `lerobotv2.1` | | --- | ---: | ---: | | `season_POC22061_2026_06_11_14_29_08_train` | 3.64 GB | 3.65 GB | | `season_POC22061_2026_06_11_19_10_57_train` | 3.70 GB | 3.71 GB | | `season_POC22061_2026_06_11_20_21_30_train` | 5.11 GB | 5.13 GB | | `season_POC22061_2026_06_14_10_25_39_train` | 4.10 GB | 4.12 GB | | `season_POC22061_2026_06_14_15_44_09_train` | 7.95 GB | 7.97 GB | | `season_POC22061_2026_06_15_10_15_11_train` | 4.37 GB | 4.38 GB | | `season_POC22061_2026_06_15_11_15_27_train` | 3.13 GB | 3.14 GB | | `season_POC22061_2026_06_15_15_56_02_train` | 10.70 GB | 10.73 GB | | `season_POC22061_2026_06_15_19_24_12_train` | 6.94 GB | 6.97 GB | | `season_POC22061_2026_06_16_10_36_10_train` | 2.65 GB | 2.66 GB | ## Usage Recommendations This release is provided as training data. For local experiments, users may split by season to avoid mixing demonstrations from the same collection session across train and evaluation sets. For policy learning, a typical setup is: - Visual observations: one or more `observation.images.*` streams - Proprioception: `observation.state` - Optional tactile signal: `observation.tactile` - Supervision target: `action` For multi-view policies, start with: ```text observation.images.head_left observation.images.head_right observation.images.wrist_left observation.images.wrist_right ``` Then add tactile video streams if your model can use high-resolution tactile observations: ```text observation.images.tactile_deform observation.images.tactile_raw ``` ## Dataset Notes - This repository contains only the LeRobot exports prepared for release. - Raw POC recording folders and intermediate HDF5 folders are intentionally excluded from this release. - The dataset is released by season, not as a single flattened LeRobot root. - Every released season includes both `lerobot3.0` and `lerobotv2.1`. - `motor_j0` to `motor_j6` are the torso/motor-related dimensions in `observation.state` and `action`. - The task is contact-rich and precision-sensitive: policies should expect object contacts, hand-part occlusions, tactile events, and fine pose adjustments. ## License and Terms Please follow the license and usage terms specified by this Hugging Face dataset repository. If you use the dataset for challenge participation, also follow the official competition rules and evaluation protocol. ## Citation If this dataset contributes to your research, please cite or acknowledge the dataset. ```bibtex @misc{roco_task_board_assembly_2026, title = {RoCo Task Board Assembly LeRobot Dataset}, howpublished = {\url{https://huggingface.co/datasets/SharpaIT/RoCo_TaskBoardAssembly}}, year = {2026} } ```