# dreamzero0/dreamzero

Code to pretrain, fine-tune, and evaluate DreamZero and run sim & real-world evals

Repository: https://github.com/dreamzero0/dreamzero
Canonical: https://ross.abutalabs.com/products/dreamzero
Homepage: https://dreamzero0.github.io/
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-04-19T06:40:39+00:00

## Health v2 (maintenance only)
Score: 50/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 78, release rhythm 35, longevity 15
- inputs: {"age_days": 218, "days_push": 136, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2593, forks 226 (observed 2026-08-28T04:07:02.993833+00:00)

## What it is
DreamZero is NVIDIA's World Action Model (WAM) that jointly predicts future video and actions from a pretrained video diffusion backbone, enabling zero-shot robot policies with real-time closed-loop control. The repository provides pretrained checkpoints, distributed WebSocket inference servers, simulation and real-robot evaluation harnesses, and LoRA/full fine-tuning code for adapting to new robot embodiments.

## Use cases
- run zero-shot robot policies on unseen tasks
- serve a video diffusion policy over websocket for real-time robot control
- fine-tune a world action model on a new robot embodiment with minutes of data
- evaluate robot policies in DROID simulation and RoboArena
- post-train DreamZero-AgiBot on play data for pick-and-place
- generate and save rollout videos from a robot policy

## When to choose
- you need state-of-the-art zero-shot or few-shot robot manipulation policies
- you want to adapt a generalist policy to a custom robot with minimal demonstration data
- you need real-time (7Hz) closed-loop control from a large video diffusion model on GB200/H100 GPUs
- you are researching world models, VLA/WAM policies, or cross-embodiment transfer

## When to avoid
- you lack high-end NVIDIA GPUs (GB200/H100) for inference or training
- you need a lightweight plug-and-play robotics library rather than a research codebase
- your task is outside robot manipulation or you have no robot/simulation setup
- you need production-supported software with long-term stability guarantees

## Facets
- artifact type: framework
- maturity: active
- function: machine-learning, deep-learning, llm-inference, video-processing, robotics, simulation
- domain: robotics, machine-learning, deep-learning, artificial-intelligence, autonomous-vehicles, simulation
- platform: python
- tags: world-action-model, video-diffusion, robot-policy, vla, zero-shot-policy, fine-tuning, websocket-inference-server, cross-embodiment-transfer, nvidia, linux, gpu, docker

## Member repositories
- dreamzero0/dreamzero (main) score 50

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:02.993833+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T02:22:16.192254+00:00, confidence not recorded.
  - readme: https://github.com/dreamzero0/dreamzero (fetched 2026-08-28T04:07:02.993833+00:00, sha 303ad84367dc)
  - homepage: https://dreamzero0.github.io/ (fetched 2026-08-29T10:04:27.038114+00:00, sha c5490991d179)
- Data as of 2026-08-30T08:39:29.467469+00:00.