Open Source RoboticsRobots, software & learning

BridgeData V2

Software

Open-access WidowX 250 robotic manipulation dataset (CC BY 4.0) paired with MIT-licensed JAX training, data-processing, and policy evaluation code. [2] [3] [4]

Installation & first result →Official websiteSource codeGetting started

For
Robot learning researchers; Offline RL and imitation-learning practitioners; VLA pretraining dataset engineers [2] [4]
Hardware
WidowX 250 6DOF robot arm (5 Hz control, VR teleoperation), Over-the-shoulder RGB-D camera, two randomized RGB… [4]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system, compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
MIT, CC-BY-4.0 · license not fully assessed for model weights · Details by part

Requirements · Research details and dates

On this page

Overview

BridgeData V2 pairs a WidowX 250 6-DoF manipulation dataset with a JAX codebase (rail-berkeley/bridge_data_v2) for goal-conditioned and language-conditioned offline policy learning. UC Berkeley RAIL publishes explicit licensing for both the data (CC BY 4.0) and the codebase (MIT). The main practical friction is choosing between the raw and pre-processed RLDS formats and deciding whether to use the repo's pinned 2023–2024 JAX 0.4.13 environment or load the RLDS split into a newer training stack. [1] [2] [3] [4]

Is this for you?

Workflows
  • Download pre-processed 256x256 TFDS/RLDS split for Octo/RLDS loaders
  • Convert raw archives via bridgedata_raw_to_numpy.py and bridgedata_numpy_to_tfrecord.py
  • Train GCBC, DDPM-BC, GCIQL, Contrastive RL, or LCBC in JAX
  • Evaluate checkpoints against WidowX 250 via bridge_data_robot

[2]

Known limitations
  • Repository pins older JAX 0.4.13 (CUDA 11 or TPU) on Python 3.10 and has had no commits since 2024-03-17.
  • Majority of episodes include only the primary over-the-shoulder camera view because additional randomized RGB and wrist cameras were added later during collection.
  • Real-robot evaluation requires the separate bridge_data_robot WidowX controller stack; pretrained checkpoint directory does not carry a separate verified license file in the inspected pages.

[1] [2] [4]

What you need

Languages / runtimes
  • Python

[2]

Dependencies
  • Python 3.10
  • jax==0.4.13 (CUDA 11 or TPU)
  • TensorFlow / tf.data / TFDS
  • bridge_data_robot (for physical WidowX evaluation)

[2]

Host computer

NVIDIA GPU (CUDA 11 for pinned jax==0.4.13) or TPU documented for training; no explicit VRAM or disk-size minima listed in README.

[2]

Skills
  • Python 3.10
  • JAX / Flax
  • TensorFlow Datasets / RLDS
  • Conda environment management

Inferred from README environment and data-loader instructions.

[2] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [2]

  • Choose between the pre-processed 256x256 TFDS/RLDS download (recommended with Octo data loader) or raw JPEG/PNG/pkl archives (demos*.zip, scripted*.zip) converted via bridgedata_raw_to_numpy.py and bridgedata_numpy_to_tfrecord.py.
  • Create a Python 3.10 conda environment, install the repo and requirements.txt, and install jax==0.4.13 for CUDA 11 or TPU.
  • Run experiments/train.py with --config experiments/configs/train_config.py:<METHOD> or evaluate downloaded checkpoints via experiments/eval.py with bridge_data_robot.

[2]

Installation

Create conda env (python=3.10), run pip install -e . and pip install -r requirements.txt, then install jax[cuda11_pip]==0.4.13 or jax[tpu]==0.4.13.

[2]

Hardware, models & data

Supported hardware
  • WidowX 250 6DOF robot arm (5 Hz control, VR teleoperation)
  • Over-the-shoulder RGB-D camera, two randomized RGB cameras, and wide-angle wrist RGB camera (640x480 raw)

[4]

Interfaces
  • tf.data custom loader (jaxrl_m/data/bridge_dataset.py)
  • TFDS / RLDS dataset format
  • CLI training/eval scripts (experiments/train.py, experiments/eval.py)
  • Server-client or Docker Compose robot controller bridge

[2]

Inputs and outputs

Raw trajectories store 640x480 RGB/RGB-D images (JPEG/PNG) and pickle state/action files at 5 Hz (38-timestep average); pre-processed TFDS/RLDS split stores 256x256 downsampled trajectories.

[2] [4]

File formats
  • JPEG/PNG + pickle (raw demos*.zip, scripted*.zip)
  • NumPy -> TFRecord (custom tf.data pipeline)
  • TFDS / RLDS (256x256 pre-processed)

[2]

Algorithms and models
  • Goal-conditioned BC (GCBC)
  • Goal-conditioned Diffusion Policy (D-GCBC / gc_ddpm_bc)
  • Goal-conditioned IQL (GCIQL)
  • Goal-conditioned Contrastive RL (CRL)
  • Language-conditioned BC (LCBC)

[2]

Evaluation

Documented real-robot evaluation on WidowX 250 via experiments/eval.py using server-client or Docker Compose against bridge_data_robot; not executed in this review.

[2]

Project notes

Raw 640×480 archives vs. pre-processed 256×256 RLDS

All demonstrations were collected on a WidowX 250 6-DoF arm teleoperated with a VR controller at 5 Hz, averaging 38 timesteps per trajectory, with raw images saved at 640×480. One camera caveat from the project page is easy to miss: extra randomized RGB views and the wide-angle wrist camera were added over time, so most trajectories contain only the primary fixed over-the-shoulder view and very few contain all four views. [4]

Upstream hosts two distinct data routes at rail.eecs.berkeley.edu/datasets/bridge_release/data/: [2]

  • Pre-processed TFDS/RLDS (tfds folder): trajectories downsampled to 256×256 in RLDS format. Upstream recommends loading this split with the Octo data loader, or regenerating custom RLDS resolutions/keys via kvablack/dlimp/tree/main/rlds_converters.
  • **Raw archives (demos*.zip for teleop, scripted*.zip for scripted pick-and-place)**: JPEGs, PNGs, and .pkl files. Training with this repo's custom tf.data loader (jaxrl_m/data/bridge_dataset.py) requires a two-stage local conversion using data_processing/bridgedata_raw_to_numpy.py followed by bridgedata_numpy_to_tfrecord.py.

Training environment and real-robot evaluation boundaries

The repository implements five baseline methods from the CoRL 2023 paper—gc_bc, gc_ddpm_bc, gc_iql, contrastive_rl_td, and language-conditioned BC (LCBC)—and links pretrained 128×128 and 256×256 checkpoints with companion JSON configs (noting that ACT and RT-1 checkpoints are not included). Setup targets Python 3.10 with jax[cuda11_pip]==0.4.13 (or jax[tpu]==0.4.13), and the repository's HEAD commit bc60a35b701a12021c8c95e9d8601274d3acd928 dates from 2024-03-17. Teams on CUDA 12+ GPUs should treat the JAX pins as an archival reference rather than a drop-in modern environment. In addition, the inspected README and project page do not specify dataset download sizes or minimum GPU VRAM requirements, so storage and training-hardware capacity must be established separately. [1] [2] [4]

Physical evaluation (experiments/eval.py) is not self-contained: it depends on installing the separate rail-berkeley/bridge_data_robot WidowX controller stack and running either a server-client bridge (recommended upstream) or Docker Compose services. [2]

What we checked

Primary-source repository metadata, README, LICENSE, and project website inspection only. No dataset archives or checkpoints were downloaded, no JAX training runs were executed, and no WidowX hardware was tested.

Not yet established

  • Requirements: operating system

Updates

Upstream last updated
Mar 17, 2024 [1]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
rail-berkeley/bridge_data_v2@bc60a35b701a12021c8c95e9d8601274d3acd928 · combined hardware and software repository

Source files and licenses

Core open-access manipulation dataset with verified CC BY 4.0 data license and MIT codebase; marked include_partial because hosted pretrained checkpoints lack a separately verified license text in the inspected pages.

LayerAvailability / license
softwareavailable · MIT
Root LICENSE in rail-berkeley/bridge_data_v2 declares MIT License. [1] [3]
model weightsavailable · License not fully assessed
README links pretrained GCBC, LCBC, D-GCBC, GCIQL, and CRL checkpoints (128x128 and 256x256) with JSON configs, but no separate checkpoint license text was inspected at that host. [2]
datasetsavailable · CC-BY-4.0
Project page explicitly states: "All data is provided under the Creative Commons Attribution 4.0 International License." [2] [4]
documentationavailable · MIT
Repository README covered by root MIT LICENSE. [2] [3]

Clear data and code licenses; check checkpoint terms separately

The bridge_data_v2 repository carries a root MIT LICENSE, and the BridgeData V2 project page explicitly states that all data is provided under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Pretrained checkpoints are linked from a separate UC Berkeley directory without an explicit checkpoint license notice in the inspected pages, so verify checkpoint redistribution terms independently if shipping weights. [2] [3] [4]

No dataset archives, RLDS shards, or checkpoints were downloaded, and no JAX training or WidowX evaluation was executed in this review. For most new workflows, the lowest-friction next step is inspecting the pre-processed 256×256 TFDS/RLDS split rather than rebuilding the legacy JAX 0.4.13 / TFRecord pipeline. [2]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history