An open-weight world action model turning video pretraining into physical control.
Today, we are releasing Praxis-1, an open-weight foundation model designed to bridge generative world simulation with physical robotic action.
While large multimodal models have achieved remarkable semantic understanding, translating perceptual video representations into precise spatial and motor actions remains a critical frontier.
Praxis-1 leverages billions of video training tokens to learn intuitive physics, spatial occlusion, and multi-object manipulation directly from vision.
From simulation to physical action
Praxis-1 transforms raw visual streams into structured spatial trajectories, enabling robotic arms and autonomous systems to operate in unpredictable real-world environments.
- Zero-Shot Tool Use: Generalizes to unfamiliar objects and kitchen utensils through learned physical affordances.
- High-Frequency Control: Generates 50Hz action trajectories while maintaining real-time perceptual feedback.
- Robust to Occlusion: Predicts object continuity when sight lines are temporarily blocked.
Benchmarks and evaluations
Across standard manipulation benchmarks including RLBench and RoboSuite, Praxis-1 sets a new state-of-the-art for open-weight visual policy models.
- 42% improvement in task completion rates over baseline vision-action transformers.
- 3x reduction in sample complexity during fine-tuning on custom enterprise robotic setups.
Model weights and developer access
Model weights, inference code, and evaluation datasets are available immediately on Hugging Face and GitHub for academic and commercial use under an open research license.
