BridgeData V2: A Dataset for Robot Learning at Scale

Source

Core Claim

BridgeData V2 is a real-robot manipulation dataset used as an important substrate for language-conditioned robot policies and robotic world-model evaluation.

Sensor-Time-Series Notes

  • The useful modeling unit is a language-conditioned manipulation trajectory with image observations and robot control inputs.
  • The dataset is especially relevant to action-conditioned latent world models because several later studies use Bridge-style rollouts to test whether generated future observations preserve action-relevant state.
  • It should be treated as visual robot trajectory data rather than as a generic numeric forecasting benchmark.

Open Questions

  • Which representation of BridgeData V2 actions is most useful for cross-dataset training: raw robot commands, normalized end-effector deltas, or learned action tokens?
  • How much proprioceptive or force/contact information is needed in addition to camera history for robust Bridge-style world models?