Perceptron’s Isaac 0.5 Targets Open Robot Control

Isaac 0.5 is a credible open-weight physical-AI bet because it trains perception, world-state estimation, reasoning, and control as one system. The catch is operational: its scale and runtime dependencies make it better suited to robotics labs than casual edge experimentation.

• –Training spans 3 trillion multimodal tokens, 1 million hours of general video, 100,000 hours of robot experience, and more than 35 robot systems. [Model card](https://huggingface.co/PerceptronAI/Isaac-0.5)
• –One backbone can produce text, coordinates, task states, and robot actions, reducing the need to stitch together separate perception and control models.
• –Sparse routing lowers per-token compute, but the full 36B checkpoint still requires substantial storage and infrastructure.
• –Continuous Flow-based control and discrete FAST action tokens give developers multiple policy interfaces.
• –The open stack has friction: direct Transformers and stock LeRobot usage are unsupported, and the repository notes additional runtime dependencies. [Repository](https://github.com/perceptron-ai-inc/isaac)