What Is BridgeData V2?

BridgeData V2 is introduced in BridgeData V2: A Dataset for Robot Learning at Scale (arXiv 2308.12952), submitted 24 August 2023, and on an official project page. The authors describe it as a large and diverse dataset of robotic manipulation behaviors designed to facilitate research on scalable robot learning, compatible with open-vocabulary, multi-task methods conditioned on goal images or natural language instructions. This is the open robot learning dataset many policies are trained and compared on, and it is third-party research.

BridgeData V2 is third-party research. Firsthand is not part of it, endorsed by it or affiliated with its authors, and makes no claim that BridgeData V2 or any other model uses or was trained on Firsthand data. The sources read describe teleoperated and scripted robot trajectories, not egocentric human video. What Firsthand states is the human side of a robot-learning mix: egocentric video, depth, 3D hand pose and action segments from paid, consented contributors, collected fast, on demand and to a written skill spec.

The abstract states 60,096 trajectories collected across 24 environments on a publicly available low-cost robot. The project page splits them into 50,365 teleoperated demonstrations and 9,731 rollouts from a scripted pick-and-place policy, and lists 13 skills. Each trajectory is labeled with a natural-language instruction for the task the robot is performing.

The 24 environments are grouped into 4 categories. Most of the data comes from 7 distinct toy kitchens with some combination of sinks, stoves and microwaves; the rest come from tabletops, standalone toy sinks, a toy laundry machine and more. Most of the data is foundational object manipulation such as pick-and-place, pushing and sweeping. Further data covers opening and closing doors and drawers, and the remainder covers stacking blocks, folding cloths and sweeping granular media.

The project page states all the data was collected on a 6-DoF robot arm, and demonstrations were collected by teleoperating it with a VR controller. The control frequency is 5 Hz and the average trajectory length is 38 timesteps.

Sensing is an RGBD camera fixed in an over-the-shoulder view, two RGB cameras whose poses are randomized during collection, and an RGB camera attached to the robot’s wrist. Images are saved at 640x480. The page adds that data collectors change the randomized views every 50 trajectories, and that cameras were added during collection, so most of the data holds only the primary fixed view and very little holds all four.

The project page says the dataset downloads as JPEGs, with the teleoperated and scripted data in separate zip files. The repository README describes the raw dataset as JPEGs, PNGs and pkl files, and offers two loaders: a custom tf.data loader, and a pre-processed TensorFlow Datasets version downsampled to 256x256, structured in the RLDS format. It recommends the Octo data loader for the RLDS version.

The project page states all data is provided under the Creative Commons Attribution 4.0 International License. The repository’s GitHub page lists the MIT License for its code.

The Open X-Embodiment entry on this site describes a dataset pooled from 60 existing datasets, and the π0 entry names Bridge v2 alongside Open X-Embodiment and DROID as open-source robot data. The sources read for this page do not mention Open X-Embodiment, so whether BridgeData V2 is among its pooled datasets is Not stated in the sources read.

The README points to the Octo data loader for the RLDS version, and lists RT-1 among the methods whose official implementations it links. The project page’s language-conditioned task videos include an RT-1 column. The abstract says the authors trained 6 imitation learning and offline reinforcement learning methods on the dataset and that performance improves with more data and higher capacity models. These are the authors’ results on their own robot, and Firsthand draws no comparison from them.

It shows the shape of open robot data: trajectories from one shared, low-cost hardware setup, each tied to a language instruction, spread across many environments. The authors say the dataset continues to grow, and the abstract credits variety of skills with improved generalization.

The VLA entry on this site says mapping language to action needs frames, instruction and action on a timeline where the three are provably simultaneous. Coverage beyond the toy kitchens, tabletops and 13 skills listed here is Not stated in the sources read.

Firsthand does not claim that egocentric human video replaces robot demonstrations or substitutes for BridgeData V2. It collects egocentric video with paired instruction and action signals on demand, to a written spec, for skills or environments not in an open dataset.

Confirm the spec on a Spec Pilot. The median from signed spec to first validated batch is 48 hours where coverage is deep and three to four weeks where it is thin. Billing is per validated hour.

What is BridgeData V2?

An open robot manipulation dataset of 60,096 trajectories across 24 environments, collected on a publicly available low-cost robot.

What does BridgeData V2 contain?

50,365 teleoperated demonstrations and 9,731 scripted pick-and-place rollouts across 13 skills, each labeled with a natural-language instruction.

How was BridgeData V2 collected?

By teleoperating a 6-DoF robot arm with a VR controller at 5 Hz, with fixed, randomized and wrist cameras.

What licence does BridgeData V2 use?

The project page states all data is under the Creative Commons Attribution 4.0 International License, and the repository page lists the MIT License for its code.

Is BridgeData V2 part of Open X-Embodiment?

Not stated in the sources read for this page. The sources read do not mention Open X-Embodiment.

Does BridgeData V2 include egocentric human video?

Not described in the sources read. The data is teleoperated and scripted robot trajectories.

Is Firsthand data part of BridgeData V2?

No. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data.