NVIDIAUnited States2026
DreamZero
An AI model: a world action model learns to predict how a scene will change and which actions lead there, so that a robot can plan its movements from what it expects to see. This page describes the software, not a robot.
VersionsDreamZero-DROIDDreamZero-AgiBot
The developer has released this model in several versions. Each figure applies to the version its source names.
Key figures as stated by the developer
5 of 6 stated · 1 gap-
Model type C
world action model - Controls not stated
-
Inputs A
cameralanguagerobot state - Access C open download
- Parameters A 14,000,000,000
- Base model A Wan2.1-I2V-14B-480P
How to read the figures
- 50HzA
- The letter in the grey square after a figure points to the source the figure came from. Pages with several sources carry several letters.
Versions and notes in the key figures (6)
Model type
Developer wording: "DreamZero is a World Action Model that jointly predicts actions and videos, achieving strong zero-shot performance on unseen tasks."
Controls
Note Model card: "Outputs: Joint autoregressive prediction of future video frames and robot actions."
Inputs
Model card: "Inputs: Visual context (via VAE), language instructions (via text encoder), and proprioceptive state." Visual context = camera, language instructions = language, proprioceptive state = robot state.
Access
README: "We release a 14B pretrained DROID checkpoint on Huggingface". The README also lists the "Released DreamZero-AgiBot checkpoint and post-training code for efficient few-shot adaptation."
Parameters
Model card for the DROID checkpoint: "Parameters: 14 Billion"
Base model
Model card: "built upon a pretrained video diffusion backbone (Wan2.1-I2V-14B-480P)"
Full specifications
9 of 13 fields stated| Field | Value | Source and caveats |
|---|---|---|
| Input and output | ||
| Model type | world action model | C Developer wording: "DreamZero is a World Action Model that jointly predicts actions and videos, achieving strong zero-shot performance on unseen tasks." |
| Inputs | camera language robot state | A Model card: "Inputs: Visual context (via VAE), language instructions (via text encoder), and proprioceptive state." Visual context = camera, language instructions = language, proprioceptive state = robot state. |
| Controls | not stated | Note Model card: "Outputs: Joint autoregressive prediction of future video frames and robot actions." A |
| Size and speed | ||
| Parameters | 14,000,000,000 | A Model card for the DROID checkpoint: "Parameters: 14 Billion" |
| Control rate | 30Hz | B Action execution rate in the paper's experiments: "at an action horizon of 48 steps at 30Hz control frequency (1.6 seconds per chunk) for bimanual manipulation robots". For the DROID checkpoint: "For DROID training, the video is sampled at 5FPS ratio and action is sampled at 15Hz." Action chunks are generated more slowly: "this 14B model is capable of real-time closed-loop control at ~7Hz". |
| Where it runs and what it is built on | ||
| On robot or cloud | not stated | Note README: the package runs inference as a server, "run distributed inference via a WebSocket server"; "Minimum: 2 GPUs for distributed inference" C |
| Base model | Wan2.1-I2V-14B-480P | A Model card: "built upon a pretrained video diffusion backbone (Wan2.1-I2V-14B-480P)" |
| Openness | ||
| Access | open download | C README: "We release a 14B pretrained DROID checkpoint on Huggingface". The README also lists the "Released DreamZero-AgiBot checkpoint and post-training code for efficient few-shot adaptation." |
| License | CC-BY-NC-4.0 | A Model card header for the weights: "License: cc-by-nc-4.0". The code repository: "This project is licensed under the Apache License 2.0". |
| Free or paid | open, non-commercial only | A Model card header for the weights: "License: cc-by-nc-4.0". |
| The developer’s own claims | ||
| Learns new tasks | without examples | A Developer claim: the DROID checkpoint is "achieving strong zero-shot performance on unseen tasks without requiring pretraining on massive, large-scale robot datasets". |
| Works across robots | not stated | Note The paper tests two robots, "We validate our main hypotheses about learning from diverse data on two robot embodiments: the AgiBot G1 mobile bimanual manipulator and the Franka single-arm robot." "We pretrain separately for each embodiment, leaving multi-embodiment training for future work." It also reports that "a model pretrained on AgiBot G1 adapts to an entirely new robot (YAM) with only 30 minutes of play data". B Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count. |
Not stated: Degrees of freedom controlled
Notes
- NVIDIA GEAR Lab describes DreamZero as a world action model that jointly predicts actions and videos.
- The DreamZero paper was submitted to arXiv on 17 February 2026.
- NVIDIA Corporation lists its address as Santa Clara, California.
Sources
-
A
huggingface.
co/ · RetrievedGEAR-Dreams/ DreamZero-DROID -
B
arxiv.
org/ · Retrievedhtml/ 2602. 15922 -
C
github.
com/ · Retrieveddreamzero0/ dreamzero
One line per source, not per figure: a source has one retrieval date. The letter beside a figure points to the line it came from.