Ex-Meta Researchers Launch Perceptron to Bring Visual AI to Industrial Robots

Artificial intelligence has largely advanced through digital applications, but a growing number of companies are working to bring AI capabilities into physical environments. Perceptron, a startup founded by two former Meta research scientists, is developing visual AI models designed to help machines better understand and interact with the real world.

Founded in November 2024, Perceptron focuses on what it describes as frontier vision models for industrial applications. The company recently introduced Isaac 0.5, an open-weight model designed to help robots perceive their surroundings, reason through tasks and take actions in environments such as warehouses and factory floors.

The model is also designed to help companies extract useful visual information from videos captured by robots. Because Isaac 0.5 is released as an open-weight model, developers and researchers can inspect its parameters and training materials.

Isaac 0.5 Targets Flexible Robot Operations

Perceptron was founded by Armen Aghajanyan and Akshat Shrivastava, who previously worked at Meta’s Fundamental AI Research (FAIR) division. The founders believe industrial automation needs more flexible AI systems that can adapt across different environments, rather than models built for a single repetitive task.

The company said many existing physical AI systems force a choice between broad foundation models that require significant computing resources and specialised models focused on limited perception or control tasks. Perceptron says Isaac 0.5 is designed as a general-purpose system that combines those capabilities.

Shrivastava explained that even a simple industrial activity, such as sorting packages, requires multiple steps. A robot must identify labels, determine the positions of objects, decide which items to pick up, and plan the sequence of actions needed to complete the task.

Perceptron’s software is intended to help robots handle those steps more flexibly. While individual technologies for tasks such as object recognition and movement control already exist, the company says fewer systems are designed to combine those abilities across changing situations.

Training Visual AI With Large-Scale Video Data

Isaac 0.5 learns from large amounts of visual information. Perceptron said the model was trained using one million hours of general video data, along with egocentric video captured from a person’s perspective while completing physical tasks and UMI video used to teach AI systems about human movements.

The company has not disclosed the specific sources of its training data. Shrivastava said Perceptron built internal petabyte-scale datasets containing multiple types of information, including images, text, video and robotic movement data.

The startup plans to offer its technology across industries including manufacturing, logistics, warehousing, security, mobility and media and entertainment.

Perceptron previously raised $16 million in 2024 from Bessemer Venture Partners, The Explorer Fund and SmartGateVC, according to PitchBook.