Perceptron Open-Sources Isaac 0.5 Robot Foundation Model

Perceptron AI has released Isaac 0.5, a 36-billion-parameter foundation model for robots, and made the weights, technical report, fine-tuning code, and inference code publicly available. The Washington-based startup is positioning the model as a general-purpose brain for embodied machines rather than a system tuned to a single robot arm or task.

Isaac 0.5 fuses video understanding, embodied reasoning, and motor control into one architecture. That combination lets a single model watch a scene, reason about what's happening, and issue physical actions, instead of stitching together separate perception and control pipelines as many robotics stacks still do.

On the LIBERO manipulation benchmark, Isaac 0.5 posted a 97.2% average success rate, edging out Nvidia's GR00T N1.7 (97.0%) and π0.5 (96.9%). The margins are thin, but the result puts an open-weight model in the same tier as leading proprietary and semi-open systems on a widely cited robotics benchmark.

Training data over hand-labeled demonstrations

The more consequential finding may be buried in the training methodology rather than the leaderboard. Perceptron says scaling general video footage from 1,000 hours to one million hours cut the amount of teleoperation data needed to reach a given performance level from roughly 5,900 hours to just 28.

Teleoperation data, where a human physically guides a robot to generate training examples, is one of the most expensive bottlenecks in robotics. If Perceptron's numbers hold up under independent testing, a model can lean far more heavily on abundant internet-scale video and far less on costly, hard-to-collect robot demonstrations.

The model was trained on three trillion multimodal tokens combining that one million hours of general video with 100,000 hours of robotics-specific experience, and it was tested across more than 35 distinct robot systems. That breadth is meant to demonstrate the model transfers across hardware rather than overfitting to one platform.

Perceptron was co-founded by CEO Armen Aghajanyan and CTO Akshat Shrivastava. By releasing Isaac 0.5's weights and code openly, the company is betting that robotics developers will build on and extend the model rather than train comparable systems from scratch, an approach that mirrors the open-weight strategy language-model labs have used to compete with closed alternatives.

For robotics teams without the budget to collect thousands of hours of teleoperation data, an open model that claims benchmark parity with GR00T and π0.5 while requiring far less demonstration data addresses a real cost problem. Whether Isaac 0.5 performs as well outside curated benchmarks, on real hardware in unstructured environments, is the next question independent developers will now be able to test for themselves.