NVIDIA / DreamZero
- joint prediction
- 2026
- Wan 2.1 (14B)
- World action model
Model cardPaper ↗
DreamZero is NVIDIA's 14-billion-parameter joint-prediction WAM: one Wan video backbone denoises future frames and actions together.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | DreamZero |
|---|---|
| Lab | NVIDIA |
| Year | 2026 |
| Paradigm | joint prediction |
| Backbone | Wan 2.1 (14B) |
| Source | https://arxiv.org/abs/2602.15922 |
What the paper claims
- DreamZero is built on Wan2.1-I2V-14B-480P. It jointly generates video and actions so that motor commands are aligned with predicted visual futures, which the authors contrast with VLAs that inherit semantic priors but not motion. Ye et al., 2026
- The authors report that larger pretrained video models produce higher-quality video predictions that translate into better action execution, and that a diverse data mix beats the same number of hours of repetitive multi-task demonstrations in their ablations. Ye et al., 2026
- System work they call DreamZero-Flash, plus parallelism, caching, quantization and kernel tuning, is reported as a 38× inference speedup, with closed-loop control at about 7 Hz. Ye et al., 2026
What it evaluates
- On AgiBot G1, trained on about 500 hours of teleoperation, they report 62.2% average task progress against 27.4% for the best pretrained VLA baseline in that table, and 39.5% versus 16.3% on their unseen-verb split. Ye et al., 2026
- On DROID-Franka they report 49% task progress and 22.5% success, against 33% / 7.5% for π0.5 and 31% / 12.5% for GR00T N1.6 in that comparison. They also report video-only transfer from humans (12 minutes) or another robot (20 minutes) improving unseen-task performance by more than 42% relative, and few-shot adaptation to a new embodiment with 30 minutes of play data. Ye et al., 2026
- Reuss's April 2026 RoboArena snapshot, which he presents as a field-map figure rather than a paper table, puts DreamZero at 1,750 Elo against 1,622 for π0.5. He notes that snapshot was trained only on DROID. That Elo pair is not in the DreamZero abstract we used; it is Reuss's reading of the public leaderboard. Reuss, NVIDIA Technical Blog (15 June 2026)
On this register
Joint prediction. Emits video and actions together. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other joint prediction models on the register: GR-1, GENIMA, Unified Video Action, Cosmos Policy.
No other register row currently lists the backbone Wan 2.1 (14B). Matching is exact; nearby names such as Wan 2.1 versus Wan 2.2 are kept separate.
Limits the authors state
- The authors say DreamZero inherits behavior-cloning limits on tasks that need sub-centimetre precision, such as key insertion or fine assembly, because their diverse pretraining underrepresents dense demonstrations for those skills. Ye et al., 2026