SKU DEM-6002 · Sold by External
BridgeData V2
Product specifications
| SKU | DEM-6002 |
|---|---|
| Data type | Demonstrations |
| Volume | 60096 trajectories |
| Size on disk | ~441 GB raw (demos_8_17.zip = 411 GB teleop + scripted_6_18.zip = 30 GB); a downsampled TFDS/RLDS version is also provided |
| Format | Per-trajectory raw files (obs_dict.pkl, policy_out.pkl, agent_data.pkl, lang.txt, images0/im_*.jpg 640x480 JPEG); also NumPy and TFRecord/RLDS (TFDS, 256x256) conversions |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open dataset (CC-BY-4.0) |
| Quality score | — |
| License | Creative Commons Attribution 4.0 International (CC BY 4.0) |
BridgeData V2 is a large-scale robot-learning demonstration dataset collected by UC Berkeley's Robotic AI & Learning (RAIL) lab using a low-cost WidowX 250 6-DOF arm. It contains 60,096 trajectories: 50,365 human-teleoperated demonstrations and 9,731 rollouts from a scripted pick-and-place policy, gathered across 24 environments (mostly 7 toy kitchens, plus tabletops, toy sinks, and a toy laundry machine) spanning 13 manipulation skills. Each trajectory is a short (~38 timestep, 5 Hz) sequence of RGB(D) images from an over-the-shoulder camera (plus randomized side views and a wrist camera), robot proprioceptive state, and 7-D end-effector actions, and is annotated with one or more crowd-sourced natural-language instructions. The data is distributed as raw JPEG/PNG + pickle files, a NumPy conversion, and a pre-processed TFDS/RLDS version, and is intended for goal-image- and language-conditioned imitation and offline RL. All data is released under CC BY 4.0 (the accompanying code repository is MIT-licensed).