Two former Meta researchers launch open-weight industrial vision model

Perceptron, a startup founded in November 2024 by former FAIR (Meta’s Fundamental AI Research) researchers Armen Aghajanyan and Akshat Shrivastava, released this week Isaac 0.5, a vision model intended to let robots “perceive, think and act” in industrial settings. The model is released as open-weight, meaning its weights and training material are publicly available for inspection, and it is aimed directly at factory floors, warehouses and any location where machines must navigate complex physical spaces.

The founders say the current problem in Physical AI is a false choice: either generic base models that require several dedicated GPUs in the cloud for each instance, or narrow models that handle perception or control but not both. Isaac 0.5 is designed to be general-purpose, not a single-task model, but a flexible intelligence layer that can read a label on a package, analyze the spatial layout of boxes, decide which to lift and in what order, and guide the robot through each step. The industry already handles most of these tasks separately; the declared novelty is the ability to perform them sequentially without custom programming for each scenario.

Behind the model are a million hours of “general video” for learning environment, appearance and scenario recognition, supplemented by extensive use of ego video—first-person footage from a body-mounted camera (typically a GoPro or wearable) of a person performing a physical task—and UMI video, which records repetitive human actions to teach motion. The company does not disclose data sources, but Shrivastava noted that internally built datasets span petabytes and cross modalities: images, text, video and robotic trajectories alike.

Releasing the model as open-weight rather than fully open-source code allows inspection of weights and training but does not necessarily grant unrestricted commercial use, a distinction well understood by the technical community. Perceptron aims to embed the software with a range of suppliers in manufacturing, logistics and warehousing, security, mobility, media and entertainment. The founders said, “Nothing like this currently exists on the market,” yet the company has not published benchmark results against existing models, detailed architecture, nor identified customers or active pilots.

Without comparative numbers, latency on specific hardware, accuracy of label detection under varying lighting, or robustness to partial occlusion, it is difficult to assess whether Isaac 0.5 truly bridges the gap between demo and live production. The open-weight release lets the community evaluate, but until independent assessments on real robots in actual warehouses are run, the promise of “perception, thinking and action” remains a statement of intent.