BridgeData V2

A roughly 60,000-trajectory multi-task manipulation dataset collected by UC Berkeley on a low-cost WidowX arm.

BridgeData V2 was released in August 2023 by UC Berkeley's Sergey Levine group and others, collected entirely on a 6-DoF WidowX 250 arm, totaling 60,096 trajectories: 50,365 are demonstrations from VR-controller teleoperation, and 9,731 come from a scripted pick-and-place policy. It covers 24 environments — a toy kitchen, a sink, a tabletop — and 13 skill categories such as pick-and-place, pushing, wiping, opening and closing drawers, and folding cloth, recorded at a 5 Hz control frequency, with every trajectory paired with a language instruction and released under a CC BY 4.0 license. Because the hardware is cheap and the data is open, it was folded into Open X-Embodiment and is a staple for pretraining and evaluating open models like Octo and OpenVLA.

ExampleSimplerEnv recreates Bridge's WidowX setup in simulation, with four tasks — put the spoon on the towel, put the carrot on the plate, stack the green block on the yellow one, and put the eggplant in the basket — designed specifically to evaluate policies trained on Bridge data.

Also called

Bridge Dataset, Bridge V2

Related

Open X-Embodiment · Trossen Robotics WidowX 250 · SimplerEnv · OpenVLA · Octo · Teleoperation

Sources

Walke et al. 2023: BridgeData V2 (arXiv 2308.12952)
BridgeData V2 项目页 (Chinese)
GitHub: simpler-env/SimplerEnv

As of

2024-01