How did DreamGen use video world models to create synthetic training data for robot policies, and what did it measure?

Partially verified: the curator re-read none of the cited sources; S10 rests on the producer's pinned reading. DreamGen (NVIDIA et al., arXiv v2, June 2025) fine-tunes an image-to-video model on a target robot, rolls it out from an initial frame plus instruction, recovers pseudo-actions with an inverse-dynamics model or latent actions, and trains a policy on these 'neural trajectories'. Reported results: on a humanoid with GR00T N1, 14 novel behaviours 11.2% -> 43.2% and 13 novel-environment tasks 0.0% -> 28.5% (policy trained only on neural trajectories plus a single-environment pick-and-place baseline); RoboCasa solely-synthetic training reaches 20.6% average over 24 tasks; policy performance rises log-linearly with number of neural trajectories. Costs and limits: 240k RoboCasa samples took 54 hours on 1500 L40 GPUs; initial frames are provided manually; the benchmark's automatic evaluator can hallucinate. Curator qualifications: Partially verified: the curator re-read none of the cited sources; S10 rests on the producer's pinned reading. Curator GET 2026-10-23 of the sources not re-read (mechanical source evidence, not semantic verification): S10: HTTP 200, bytes differ from the producer pin. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/abs/2505.12705v2",
      "note": "candidate evidence excerpts"
    }
  ]
}