Two former Meta AI researchers, Armen Aghajanyan and Akshat Shrivastava, founded Perceptron in November 2024 to develop advanced visual AI models for industrial applications. Their latest release, Isaac 0.5, aims to equip vision-guided robots with the ability to perceive, reason, and act within complex environments like factory floors and warehouses. Unlike specialized AI tools, this model is designed as a flexible, general-purpose system capable of adapting to various tasks and scenarios in industrial automation.

Isaac 0.5 has been trained on an enormous dataset composed of over one million hours of general video footage, as well as egocentric and robotic motion data, which helps it learn to recognize environments and perform multi-step tasks such as sorting packages. Perceptron is transparent with Isaac 0.5’s development, offering the model and training details openly for inspection. This approach reflects the company’s goal to advance visual intelligence that blends perception with control, overcoming limitations of current AI systems in robotics.

The startup has already secured $16 million in funding from investors including Bessemer Venture Partners and The Explorer Fund and is in the process of closing an additional round. The founders envision their technology playing a central role in automation across manufacturing, logistics, security, mobility, and even media. They believe the increased flexibility and broader capabilities of their model distinguish it from existing AI offerings tailored to narrow, repetitive functions on the factory floor.

Perceptron’s work comes amid growing investor interest in physical AI, which is striving to bring the transformative power of AI beyond digital realms into real-world robotics. While several companies are developing task-specific robots actively deployed in industry, general-purpose robots powered by AI remain largely experimental. Experts see the current time as an early phase in physical AI development, requiring more diverse data and compute resources before robots can reliably perform a wide range of manipulation tasks autonomously.