Unified Video Action
- joint prediction
- 2025
- World action model
Model cardPaper ↗
Unified Video Action (UVA) learns one latent for video and actions, then decodes each with its own lightweight diffusion head so the policy can skip video generation at inference.
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 | Unified Video Action |
|---|---|
| Lab | — |
| Year | 2025 |
| Paradigm | joint prediction |
| Backbone | — |
| Source | https://arxiv.org/abs/2503.00200 |
What the paper claims
- UVA jointly optimizes video and action prediction through a shared latent representation, then decouples decoding so action inference can bypass video generation. Li, Gao, Sadigh & Song, 2025
- Masked training lets the same weights act as a robot policy, a video model, or a forward or inverse dynamics model, depending on which of video and action are masked. Li, Gao, Sadigh & Song, 2025
- The authors contrast this with hierarchical generate-video-then-act pipelines, which they say are slower and propagate video errors into actions. Li, Gao, Sadigh & Song, 2025
What it evaluates
- They evaluate on seven public benchmarks and report particularly strong multi-task results: 20 percentage points above the best baseline on PushT Multitask, and 5 points on LIBERO-10. Li, Gao, Sadigh & Song, 2025
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, Cosmos Policy, DreamZero.
The backbone field is unconfirmed, so this card does not claim a video-model family.
Limits the authors state
- The authors state that the framework does not currently use large amounts of actionless video, and that on some real-world tasks it is only comparable to a Diffusion Policy–UMI baseline. They leave web-scale video pretraining to future work. Li, Gao, Sadigh & Song, 2025
Coverage
No stories on this model yet.