BridgeData V2

Last updated: Oct 7, 2026

On this page

Dimensions Data composition Metrics Leaderboard How to run

BridgeData V2 (UC Berkeley RAIL) is a real-robot manipulation dataset for open-vocabulary, multi-task policies. Official numbers: 60,096 trajectories on a WidowX 250 6-DoF arm across 24 environments (13 skills) — 50,365 VR teleop demos and 9,731 scripted pick-and-place rollouts. Seven toy kitchens dominate; the rest are tabletops, sinks, and similar. Control is 5 Hz (mean length ~38 steps). Sensing is 640×480: fixed RGB-D over-the-shoulder, two randomized RGB views, and a wrist camera. Every trajectory has a language label; eval is language-conditioned (eval_lc) or goal-image-conditioned (eval_gc).

Paper: BridgeData V2. Project: rail-berkeley.github.io/bridgedata. Code: rail-berkeley/bridge_data_v2 (MIT). Official media are HTTP / project-page downloads, not a single Hugging Face table.

WorldFoundry imports policy-evaluation exports (task / goal success, offline action accuracy, rollout success). It does not materialize the dataset, run WidowX inference, or execute real-robot rollouts. Stage data under ${WORLDFOUNDRY_DATA_DIR}/datasets/bridgedata-v2 if you need it locally.

Official paper teaser: BridgeData V2’s large-scale real-robot manipulation dataset.

Dimensions

Language-conditioned eval
KindTrack
Open-vocabulary instructions on the WidowX 250. Official eval_lc path.
Goal-image-conditioned eval
KindTrack
Goal-image conditioning (official eval_gc). Distinct from language success.
Teleop vs scripted pick-and-place
KindSplit
50,365 VR-teleoperated demonstrations and 9,731 scripted pick-and-place rollouts.
13 skills / 24 environments
KindCategory
Pick-and-place, push, sweep, stack, fold, and related skills across toy kitchens, sinks, and tabletops.

Data composition

Size
UC Berkeley RAIL real-robot set: 60,096 WidowX 250 6-DoF trajectories in 24 environments (7 toy kitchens plus tabletops, sinks, laundry). 5 Hz, ~38-step mean length, 640×480 RGB-D / wrist / randomized third-person views, every trajectory language-labeled. Official download is HTTP / project-page, not a single Hugging Face table.
Sources
github.com
Splits
eval
Modalities
language · vision · action · state · proprioception
License
MIT
Notes
Code repo is MIT. Dataset materialization follows the project-page HTTP release. Hugging Face copies in the wild are third-party conversions.

Metrics

Metric

success_rate
DirectionHigher is better
Primary task success over imported policy evaluations.
PrimarySuccess Rate
task_success
DirectionHigher is better
Per-task success aggregate.
goal_success
DirectionHigher is better
Goal-image-conditioned task success (eval_gc).
action_accuracy
DirectionHigher is better
Offline action-prediction accuracy against reference actions.
rollout_success
DirectionHigher is better
Success over executed real-robot or official scripted rollouts.

Leaderboard

Results from the paper (not a live board; not WorldFoundry-run scores). Author comparison of BridgeData V2 vs V1+PTR on the paper’s two evaluated tasks. BridgeData V2: A Dataset for Robot Learning at Scale (2023).

Model	Average success (2 reported tasks)
BridgeData V2	0.70
BridgeData V1 + PTR	0.08

How to run

worldfoundry-eval zoo benchmark-run \
  --benchmark-id bridgedata-v2 \
  --mode official-validation \
  --official-results-path <official_results.json> \
  --output-dir tmp/bridgedata-v2/official-validation \
  --json

Default import path: WORLDFOUNDRY_BRIDGEDATA_V2_RESULTS_PATH. Manifest: worldfoundry/data/benchmarks/tasks/external/bridgedata-v2.yaml. Declare language vs goal-image (and teleop vs scripted) with the export. Outputs: scorecard.json (primary success_rate; also goal_success, action_accuracy, rollout_success) and raw_results.jsonl.

← Benchmark Hub