#25 · Primary category： Robotics & Embodied AI Simulation
dreamzero
Code to pretrain, fine-tune, and evaluate DreamZero and run sim & real-world evals
Project last updated：04/19/26
GitHub Stars
2.6K
Forks
229
Contributors
7
License
Apache-2.0
Why we included this project
DreamZero comes out of NVIDIA's GEAR Lab with an unusual bet: instead of training a robot policy that only outputs actions, the model predicts video frames and actions together. That choice appears to pay off. The DROID checkpoint, trained from scratch with no pretraining on large-scale robot data, tops the MolmoSpaces and RoboArena leaderboards and handles manipulation tasks it never saw during training. The repo ships that checkpoint, an AgiBot checkpoint, a distributed WebSocket inference server aimed at GB200 and H100 hardware, and fine-tuning code. The practical part for most teams is post-training: roughly 30 minutes of play data on a new robot can produce basic language-following and pick-and-place behavior, with a step-by-step guide for adding a new embodiment. Between the DROID simulation harness and RoboArena integration, it's a useful reference for anyone training or evaluating video-backbone robot policies.
Articles for this project
No articles for this project yet.
To suggest a topic or contribute an article, contact us.
Related projects in this category
genesis-world
Simulation platform for general-purpose robotics & embodied AI learning.
lerobot
🤗 LeRobot: Making AI for Robotics more accessible with end-to-end learning
ml-agents
The Unity Machine Learning Agents Toolkit (ML-Agents) is an open-source project that enables games and simulations to serve as environments for training intelligent agents using deep reinforcement learning and imitation learning.
AirSim
Open source simulator for autonomous vehicles built on Unreal Engine / Unity, from Microsoft AI & Research
carla
Open-source simulator for autonomous driving research.