mimic-video
- inverse dynamics
- 2025
- Cosmos-Predict2
- World action model
mimic-video pairs a frozen internet video model with a flow-matching inverse-dynamics decoder that reads actions from the video model's latent plan.
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 | mimic-video |
|---|---|
| Lab | — |
| Year | 2025 |
| Paradigm | inverse dynamics |
| Backbone | Cosmos-Predict2 |
| Source | https://arxiv.org/abs/2512.15692 |
What the paper claims
- The authors call this a Video-Action Model (VAM): a pretrained video diffusion backbone predicts a future trajectory in latent space, and a flow-matching decoder recovers low-level actions from those latents without full RGB generation at every step. Pai et al., 2025
- They instantiate the backbone as Cosmos-Predict2 and keep it frozen, so the video model is not trained on scarce robot action labels. The decoder is trained as an inverse-dynamics model. Pai et al., 2025
- They report 10× sample efficiency and 2× faster convergence compared with traditional VLA architectures in their comparison. Pai et al., 2025
What it evaluates
- They evaluate across single-arm and bimanual dexterous embodiments in simulation and real-world environments, and report state-of-the-art results in that comparison. Pai et al., 2025
- An oracle study in the paper conditions the decoder on ground-truth future video latents and reports near-perfect success, which they read as evidence that control quality tracks video-prediction quality. Pai et al., 2025
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, LingBot-VA, DVA.
No other register row currently lists the backbone Cosmos-Predict2. Matching is exact; nearby names such as Wan 2.1 versus Wan 2.2 are kept separate.