Rhoda AI / DVA
- inverse dynamics
- 2026
- Causal video model
- World action model
DVA is Rhoda AI's direct video-action model: a causal video generator decides the future, and a small inverse-dynamics model translates that video into motor commands.
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 | DVA |
|---|---|
| Lab | Rhoda AI |
| Year | 2026 |
| Paradigm | inverse dynamics |
| Backbone | Causal video model |
| Source | https://www.rhoda.ai/research/direct-video-action |
What the paper claims
- Rhoda defines a Direct Video-Action Model as a policy that translates predictions from a pretrained causal video model into actions in a real-time closed loop, with the video model responsible for the decision. Rhoda AI Research
- They pretrain a causal video model from scratch on web video, rather than distilling a bidirectional generator. A Context Amortization objective predicts the future at every position along a long noise-free history. An inverse-dynamics model then maps predicted video to end-effector motion. Rhoda AI Research
- They claim the inverse-dynamics translator can be trained on as little as about 10 hours of data per embodiment type, including random motions rather than task demonstrations, and that Leapfrog Inference overlaps the next prediction with the actions already in flight. Rhoda AI Research
What it evaluates
- The lab page reports data-efficient long-horizon task learning with about 10 hours of total robot data, hundreds of frames of visual context, and one-shot imitation of a human demonstration at test time. It does not publish a public simulation-benchmark table; we do not add scores from elsewhere. Rhoda AI Research
On this register
Inverse dynamics. Generates the future video, then infers the actions that would produce it. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other inverse dynamics models on the register: UniPi, Video Prediction Policy, mimic-video, LingBot-VA.
No other register row currently lists the backbone Causal video model. Matching is exact; nearby names such as Wan 2.1 versus Wan 2.2 are kept separate.
Coverage
No stories on this model yet.