DreamZero
NVIDIA, United States 路 February 2026
NVIDIA (GEAR Lab)
World model 路 Open weights, CC BY-NC 路 Also written DreamZero-DROID, DreamZero-AgiBot, World Action Models are Zero-shot Policies
- Parameters
- 14 B
- 14B according to the paper.
- GPU memory
- 26 GB
- weights at bf16, computed. Fits a 40 GB card
- Inference latency
- 150 ms
- two NVIDIA GB200 GPUs, DreamZero-Flash with one denoising step
- Weights
- CC BY-NC
- CC BY-NC 4.0 for the DreamZero-DROID checkpoint.
What it is
DreamZero is NVIDIA's World Action Model: a 14B autoregressive video diffusion model, built on Wan2.1-I2V-14B, that predicts future video frames and robot actions together, so the policy is a video world model rather than a vision language model with an action head. On an AgiBot robot pretrained on about 500 hours of teleoperation it reached 62.2 percent average task progress on seen tasks against 27.4 percent for the best pretrained VLA baseline, and on unseen DROID tasks 49 percent progress and 22.5 percent success against 31 and 12.5 percent for GR00T N1.6 and 33 and 7.5 percent for 蟺0.5. With caching and a one-step variant it runs closed loop at about 7 Hz, but on two GB200 GPUs; the released inference server needs at least two GPUs and takes about 3 s per inference on H100. A DROID checkpoint under a non-commercial licence and an AgiBot checkpoint for post-training under Apache-2.0 are public.
Architecture
- Backbone
- Wan2.1-I2V-14B-480P image to video diffusion model, run autoregressively. The paper also ablates a smaller 5B variant.
- Action head
- Joint video and action diffusion: future frames and an action chunk are denoised together, with 4 denoising steps by default and 1 step in DreamZero-Flash. Chunks of 48 actions at 30 Hz on AgiBot and 24 actions at 15 Hz on DROID.
- Action chunk
- 24 steps per forward pass
- Control rate
- 15 Hz
- Parameters
- 14B according to the paper. Hugging Face counts about 22.9B parameters in the released checkpoint files. The README puts the AgiBot checkpoint download at about 45 GB. Chunk length and rate above are for the DROID checkpoint.
- Pretraining data
- AgiBot model: about 500 hours of teleoperation data across 22 environments, 7.2K episodes of about 4.4 minutes and 42 subtasks on average. DROID model: trained on the public DROID Franka dataset alone, without robot pretraining, according to the README. The video prior comes from Wan2.1.
- Embodiments
- AgiBot robot, Franka (DROID setup), YAM arm (post-trained from 30 minutes of play data)
What hardware it needs
This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.
| Precision | Weights | Note |
|---|---|---|
| fp32 | 52 GB | Training master weights, rarely used for inference |
| bf16 or fp16 | 26 GB | The usual way these checkpoints are served |
| int8 | 13 GB | Quantised, expect some loss of precision on fine motions |
| int4 | 6.5 GB | Aggressive quantisation, verify success rate before trusting it |
Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 40 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.
| Inference latency, reported | 150 ms on two NVIDIA GB200 GPUs, DreamZero-Flash with one denoising step source |
| Control rate | 15 Hz |
The README requires at least two GPUs for distributed inference, tested on GB200 and H100, with CUDA 12.9 or newer. After warm-up one inference takes about 0.6 s on GB200 and about 3 s on H100. The paper's 150 ms (DreamZero-Flash) and about 350 ms (4 denoising steps) were measured on two GB200 GPUs.
Reported results
Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.
Simulation
No results in this category are published for this model.
Real world
Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.
- Real robot, AgiBot (DreamZero paper protocol) seen tasks62.2%average task progressAuthors' own evaluation. Best pretrained VLA baseline 27.4 percent; VLAs trained from scratch score near zero.source
Fine tuned tasks
What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.
- Real robot, DROID Franka (DreamZero paper protocol) unseen tasks49%
- Real robot, AgiBot (DreamZero paper protocol) unseen tasks39.5%
- Real robot, DROID Franka (DreamZero paper protocol) unseen tasks22.5%
On real hardware
AgiBot: 62.2 percent average task progress on seen tasks and 39.5 percent on unseen tasks, against 27.4 and 16.3 percent for the best pretrained VLA. DROID unseen tasks: 49 percent progress and 22.5 percent success. Post-training on table bussing reached 83 percent with 4 denoising steps and 74 percent with DreamZero-Flash at one step. A new arm (YAM) was adapted with 30 minutes of play data, 55 trajectories.
Fine tuning it yourself
The repository has LoRA and full fine-tuning scripts and a guide for new embodiments. Start from the DreamZero-AgiBot checkpoint; the authors report basic language following and pick-and-place after about 30 minutes of play data on a new robot. Transfer from 10 to 20 minutes of video-only demonstrations raised unseen-task performance from 38.3 to 54.3 percent with human video and to 55.4 percent with video of another robot.
Where it helps, where it does not
Strengths
- Generalizes to new tasks better than the VLAs it was compared with: 49 against 31 to 33 percent task progress on unseen DROID tasks.
- Learns from video without action labels: human or other-robot video improved unseen-task performance from 38.3 to 54.3 or 55.4 percent.
- Open checkpoints and full training code, including a DROID model trained without large-scale robot pretraining.
- Fast for its size after optimisation: 150 ms per action chunk with DreamZero-Flash, a 38 times speedup over the naive implementation.
Limits
- Hardware: the released server needs at least two GPUs, real-time control was shown on two GB200 GPUs, and one inference on H100 takes about 3 s.
- The DROID checkpoint is non-commercial (CC BY-NC 4.0).
- Results are task progress on the authors' own setups; success on unseen DROID tasks stays low at 22.5 percent.
- The one-step DreamZero-Flash costs accuracy: 74 against 83 percent on table bussing.
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://arxiv.org/abs/2602.15922
- https://arxiv.org/html/2602.15922v1
- https://dreamzero0.github.io/
- https://github.com/dreamzero0/dreamzero
- https://huggingface.co/GEAR-Dreams/DreamZero-DROID
- https://huggingface.co/GEAR-Dreams/DreamZero-AgiBot
Entry last checked 2026-11-12.
Try a policy on a real arm
A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.