NVIDIA / Cosmos Policy
- joint prediction
- 2026
- Cosmos
- World action model
Cosmos Policy fine-tunes NVIDIA's Cosmos-Predict2 video model to emit actions, future states and values as extra latent frames, with no new action architecture.
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 | Cosmos Policy |
|---|---|
| Lab | NVIDIA |
| Year | 2026 |
| Paradigm | joint prediction |
| Backbone | Cosmos |
| Source | https://arxiv.org/abs/2601.16163 |
What the paper claims
- The method adapts Cosmos-Predict2-2B through a single stage of post-training on robot demonstrations, with no architectural modifications. Actions, proprioception, future images and state values are encoded as latent frames inside the video diffusion process. Kim et al., 2026
- At test time it can act as a direct policy or plan with best-of-N sampling: generate candidate action chunks, imagine future states, rank them by predicted value, and execute the highest-value chunk. Kim et al., 2026
What it evaluates
- As a direct policy the authors report 98.5% average success on LIBERO and 67.1% on RoboCasa, and 93.6% average success on their real-world bimanual manipulation tasks, against diffusion policies, other video-based policies and fine-tuned VLAs trained on the same demonstrations. Kim et al., 2026
- With model-based planning after incorporating policy-rollout data, they report a 12.5 percentage-point higher task-completion rate on average on two challenging real-world manipulation tasks. Kim et al., 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, DreamZero.
No other register row currently lists the backbone Cosmos. Matching is exact; nearby names such as Wan 2.1 versus Wan 2.2 are kept separate.
Limits the authors state
- The authors report that model-based planning is slow (around 5 seconds per action chunk), which they say may limit dynamic tasks; that effective planning needs substantial rollout data beyond the demonstration distribution; and that they only studied best-of-N search with one layer in the search tree. Kim et al., 2026