Isaac 0.5

Isaac 0.5 is an open foundation model for robot learning developed by Perceptron AI (Perceptron, Inc.). It unites multimodal video understanding, embodied reasoning, spatial grounding, task-progress estimation and robot control in a single 36-billion-parameter sparse model. It reads images, video, language instructions, robot state and previous actions, and produces text, normalized coordinates, task-state outputs or robot actions.

Architecture

Isaac 0.5 uses a 36B Qwen-family vision-language backbone with sparse experts (mixture-of-experts). Each MoE layer gives every token access to 256 learned experts and a null route — a token may use zero to eight routed experts, while a shared expert and residual path remain active. Continuous control is handled by a dedicated Flow expert and a 36-block diffusion transformer that generates an action chunk; discrete control uses a separate vocabulary of 2,048 FAST action tokens.

Training

The model was trained on more than 35 robot systems, 100,000 hours of robot experience, one million hours of general video and three trillion multimodal tokens. Video understanding, spatial grounding, task progress, future-percept prediction and robot action are co-trained from the beginning on one shared backbone. Perceptron also reports a scaling law describing how broad video trades off against expensive teleoperation.

Availability

Model weights are published on Hugging Face under the Apache 2.0 license, while training and inference code, LeRobot integration, a reference policy server and evaluation tools live in the Perceptron Isaac repository on GitHub. The checkpoint is consumed through the Perceptron Isaac repository; direct stock Transformers or LeRobot usage is not currently supported.