NVIDIA

NVIDIA / DreamDojo

Public

Source code of DreamDojo by the NVIDIA GEAR Team.

491
32
100% credibility
Found Feb 20, 2026 at 87 stars 6x -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

DreamDojo is NVIDIA's open-source framework for training interactive world models on massive human video datasets to predict robot actions.

How It Works

1
🔍 Discover DreamDojo

You stumble upon this exciting NVIDIA project through their website or research paper, promising robot brains trained on hours of real human videos.

2
🛠️ Set up your playground

Follow simple guides to prepare your computer workspace so everything runs smoothly.

3
📥 Grab ready-made brains

Download the pre-trained models and video collections from a trusted sharing site.

4
🔧 Choose your training recipe

Pick from easy setups for different robot sizes and start teaching the model with human action videos.

5
🤖 Watch robots come alive

Your model learns to predict smooth, realistic movements from human examples, adapting to new robot friends effortlessly.

6
🧪 Test and play

Feed in robot actions and see lifelike video predictions unfold in real-time.

✨ Lifelike robot world unlocked

You now have a powerful tool generating endless robot scenarios from human videos, ready for your robotics adventures!

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 87 to 491 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is DreamDojo?

DreamDojo is NVIDIA GEAR team's Python codebase on GitHub for building interactive world models from massive human video datasets—44k hours of egocentric footage—to simulate robot behaviors. It handles pretraining on broad human data, post-training on robot-specific clips like GR1 teleop, and distillation for real-time 10 FPS autoregressive video generation conditioned on actions. Developers get pretrained 2B/14B models on Hugging Face, configs for robots like Agibot or Yam, and eval scripts to test generalization across environments.

Why is it gaining traction?

Unlike generic video diffusion models, DreamDojo excels at action-conditioned robot world modeling with strong zero-shot generalization to new objects/scenes, backed by NVIDIA's release of checkpoints, datasets, and a clear GitHub readme with setup docs. The distillation pipeline enables stable long-horizon rollouts over a minute, making it practical for robotics sims without endless compute. Python users appreciate the CLI launch scripts, Docker support, and integration with PyTorch for quick experiments.

Who should use this?

Robotics engineers fine-tuning world models on custom humanoid data, like GR1 or G1 teleop episodes. AI researchers prototyping action-to-video pipelines for planning or imitation learning. Teams at startups or labs needing fast inference for robot policy eval, especially if you're already using NVIDIA GPUs and GitHub repos for code sharing.

Verdict

Grab it if you're in robot AI—NVIDIA's polish shines through despite low 1.0% credibility from 49 stars and early-stage docs. Maturity is nascent (fresh arXiv paper, partial releases), but pretrained weights and evals make it instantly usable; star the GitHub repository and watch for distillation/teleop code.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.