Perceptron AI Launches Isaac 0.5 Open-Weight Robotics Model

Perceptron AI has launched Isaac 0.5, a 36-billion-parameter open-weight embodied foundation model that integrates video understanding, embodied reasoning, and robot control into a single unified system.

This marks the first open model at the frontier of all three capabilities, enabling industrial automation and robotics teams to deploy advanced AI for perception and control across manufacturing, logistics, warehousing, security, and mobility applications.

The model was trained on three trillion multimodal tokens, including one million hours of general video and 100,000 hours of robotics-oriented experience across more than 35 robot systems. Perceptron also established a new scaling law demonstrating that video-heavy training can reduce teleoperation data requirements by approximately 210x.

Armen Aghajanyan, co-founder and CEO of Perceptron AI, stated:

"Companies need a model that performs at the frontier, learns a new task quickly and adapts to their hardware. Isaac gives them a strong, open starting point, and our team is working alongside our customers to bring it into real operations."

On the LIBERO benchmark for robot manipulation, Isaac averages 97.2% success across spatial, object, goal, and long-horizon tasks, outperforming NVIDIA GR00T N1.7 at 97.0% and Physical Intelligence π0.5 at 96.9%. The model also learns new tasks faster, reducing error by 7.0x to 10.5x after a single expert demonstration, significantly exceeding competing models.

Akshat Shrivastava, co-founder and CTO of Perceptron AI, stated:

"A real robot workflow rarely begins and ends with one motion. The system has to understand what it sees, decide what matters and connect that decision to action. Isaac was built to carry that workflow from video and language through to control."

Perceptron works directly with customers to adapt Isaac to their cameras, robot hardware, demonstration data, and operating workflows. Customers can begin with the open model, fine-tune it on their own demonstrations, and collaborate with Perceptron on final deployment.

Headquartered in Bellevue, Washington, Perceptron AI was founded in November 2024 by former Facebook AI Research (FAIR) scientists Armen Aghajanyan and Akshat Shrivastava. The company has released model weights, technical reports, and fine-tuning tools to support commercial adoption.