Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Flash-Chord video training data

Private dataset bundle for the 2d_v2a reproduction workflow in Flash-Chord-Tactile. It contains the original action-free RGB training shards and their complete recording indexes. These are PyTorch .pt files, not MP4 previews or Hugging Face viewer tables. No model checkpoints, credentials, online action replay buffers or raw real-robot action tables are included.

Contents

Directory Episodes Shards Shard bytes Source
simulation/video_lift/ 7424 15 12762372716 RoboCasa G1 + Dex3 bottle lift: geometry demonstrator and random actions
real/video_real_pickbottle/ 202 202 818067834 Converted Unitree G1_Dex3_PickBottle head-camera video

Both directories include index.json with status COMPLETE, the exact shard list, split/source labels, frame rate, view order and recording metadata. All indexed episodes are retained, including unsuccessful simulated demonstrations. Existing index paths refer to the original research workspace for provenance.

Simulation data

Source run: demonstration_videos_bottle_lift_workspace_cameras_pressed_palm_shortest_reach. Recipe: the identically named YAML under robotic_grounding/video_to_action/configs/ in the source repository.

Split Demonstrator Random actions Scenario seeds
Train 4096 2048 1000–1011
Validation 512 256 2000–2001
Test 512 0 3000

frames is uint8 [episodes,time,views,96,96,3], with views ego_workspace then exo_workspace_right. Sampling is 5 Hz (0.2 s), with 31 frames including reset and terminal observations for each 6 s episode. Shards also preserve measured joint positions/velocities in rad and rad/s, exact object positions in simulator-world metres, mass in kg, timestamps in seconds from reset, seeds and outcomes. The video planner consumes RGB; auxiliary state is for provenance and diagnostics. No demonstration action labels are supplied to the action model.

The spatial autoencoder trains on both training sources, and the video predictor on demonstrator training episodes. The action model instead collects its own simulated transitions online. This bundle does not replace the robot/scene/object assets or the independently trained model weights.

Real robot video

Source run and conversion recipe: video_real_pickbottle / video_real_pickbottle.yaml. Original source: Unitree G1_Dex3_PickBottle_Dataset, 202 teleoperated episodes. Conversion uses only observation.images.cam_left_high from the original 30 Hz stereo video, samples at 5 Hz, centre-crops to square and resizes to 96 pixels. The converted source label is real_robot and view is cam_left_high. Episode indices 0–161 are train, 162–181 validation, and 182–201 test. There is one shard per episode; sequence lengths remain variable. These converted shards contain no robot-state/action/outcome labels. This branch supports offline video prediction; it does not demonstrate physical robot control.

Download and use

Authenticate to the account that can access this private dataset, then:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="Railgun526/flash-chord-video-data",
    repo_type="dataset",
    local_dir="/path/to/flash_chord_video_data",
)

Set custom.video_planner.data_directory in the simulation planner recipe to /path/to/flash_chord_video_data/simulation/video_lift (episode sources demonstrator and random_actions). For the real-video planner set it to /path/to/flash_chord_video_data/real/video_real_pickbottle (episode source real_robot). The existing loader reads each directory's index and all listed shards directly.

Follow the repository root README and package reproduction commands for asset preparation, planner training, online action learning and evaluation. The reference policy's performance is a simulated closed-loop result; these recordings alone are not an action-model checkpoint or a controller evaluation.

Downloads last month
17