Washington— Perceptron AI, a Washington-based robotics AI startup, has launched Isaac 0.5, a new AI model designed to help robots understand their surroundings, follow instructions and decide how to carry out tasks.

According to the company, the 36-billion-parameter model processes multiple types of information, including images, videos, written instructions, and data about robot movements. By combining these inputs, Isaac 0.5 can help robots identify objects, understand tasks, track their progress, and determine what actions to take. Perceptron describes it as the first open model to combine video understanding, embodied reasoning and robot control at the frontier level of all three.

Perceptron said in a public statement that Isaac 0.5 was trained using data from more than 35 robot systems and over 100,000 hours of robot experience. Its training also included one million hours of general video and three trillion multimodal tokens.

The company reported that scaling general video from 1,000 hours to one million cut the teleoperation needed to reach the same, well-calibrated action loss from roughly 5,900 hours to 28, according to a BusinessWire release. On the LIBERO robot-manipulation benchmark, Perceptron said Isaac achieved a 97.2% average success rate, which the company says places it ahead of other open robot models, including Physical Intelligence's π0.5 and NVIDIA's GR00T N1.7, per reporting from The AI Insider.

The company is releasing Isaac 0.5 as an open-weight model, allowing developers and researchers to access, modify and build on the system rather than relying entirely on a closed AI platform. Perceptron has released the model's weights, technical report, and fine-tuning and inference code, giving developers tools to train, adapt, and evaluate the model for different robotics applications.

Perceptron AI was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, who were previously both researchers in Meta's AI research group. Aghajanyan, the company's co-founder and CEO, said in a statement that companies need a model that performs at the frontier, learns a new task quickly and adapts to their hardware, and that Isaac gives them a strong, open starting point while the team works alongside customers to bring it into real operations, according to coverage from AIwire.

The company is positioning Isaac 0.5 as a foundation model for robots that need to do more than simply recognize objects or execute pre-programmed commands, combining perception, reasoning and physical action. Perceptron says the model has applications across manufacturing, logistics, warehousing, security and mobility, including factories and warehouses where robots may need to work alongside people and respond to changing conditions.