DreamZero
$ 0
DreamZero is an open world action model from NVIDIA’s GEAR Lab. It predicts video and robot actions together, which helps robots handle new tasks and places without extra training.
Available on backorder
Brain Score
2
Specifications and details:
| Nationality | US |
|---|---|
| Website | https://www.nvidia.com |
| Model type | Foundation Model |
| Manufacturer | NVIDIA |
| Release date | February 17, 2026 |
Description
DreamZero is a world action model from NVIDIA’s GEAR Lab. It builds on a large video diffusion model and predicts future video frames and robot actions together. The team argues that this lets a robot learn real physics, rather than just matching words to motions. As a result, DreamZero handles new tasks and places much better than typical VLA models. The project stands out as a strong case for using video models as robot brains.
In tests, the model completed far more of each unseen task than leading VLA baselines. It also learns from video of humans or other robots, so short clips can boost results on new tasks. For example, it adapted to the YAM robot with only 30 minutes of play data. Thanks to speed work, the 14-billion-parameter model runs at about seven actions per second. NVIDIA released the code under Apache 2.0, and the checkpoints are on Hugging Face. Researchers can then fine-tune it for their own robots with the included guide.
New! 2026 Humanoid
Robot Market Report
198 pages of exclusive insight from global robotics experts — uncover funding trends, technology challenges, leading manufacturers, supply chain shifts, and surveys and forecasts on future humanoid applications.
now Google DeepMind
Contact Humanoid.guide
Website: https://dreamzero0.github.io
