Ex-meta researchers launch vision AI model for industrial automation

A startup founded by former Meta researchers, Perceptron, has launched Isaac 0.5, a vision model designed to enhance the capabilities of machines in industrial environments. The model, developed by Armen Aghajanyan and Akshat Shrivastava, who previously worked at Meta’s Fundamental AI Research (FAIR) division, aims to provide robots with the ability to perceive, reason, and act in complex settings such as warehouses and factory floors. The software is also released as an open-weight model, allowing public inspection of its parameters and training materials. Isaac 0.5 is positioned as a general-purpose tool, distinguishing it from existing models that are often specialized for specific tasks. Aghajanyan and Shrivastava emphasized that their model is flexible, adapting to different environments and situations. In an interview, Shrivastava highlighted the complexity of tasks like sorting packages, which involve multiple steps including label reading, spatial analysis, and decision-making. The model is trained on extensive video data, including general video, ego video captured from a person’s perspective, and UMI video that records repetitive human actions. Perceptron’s datasets are described as petabyte-scale, spanning images, text, video, and robotic trajectories. The startup claims to have internally built these datasets, though the exact sources of the training data remain undisclosed. Perceptron aims to position itself as a leader in the wave of industrial automation, targeting industries such as manufacturing, logistics, security, mobility, and media. Aghajanyan noted that nothing similar exists currently, expressing excitement about the potential of their technology. The company previously raised $16 million in 2024 from Bessemer Venture Partners, The Explorer Fund, and SmartGateVC, and is currently in the process of closing an additional funding round.