Key facts
- Perceptron, founded by ex-Meta AI scientists, launched Isaac 0.5, an AI model for industrial automation.
- Isaac 0.5 enables robots to perceive, reason, and act in complex environments like warehouses and factories.
- The model is open-weight, allowing public inspection of its parameters and training data.
- Perceptron secured $21 million in funding led by Bessemer Venture Partners.
- The AI is designed to be general-purpose and flexible, unlike existing narrow industrial AI models.
Perceptron, a startup founded by two former Meta AI research scientists, has launched its latest AI model, Isaac 0.5. The model is designed to bring advanced visual AI capabilities to industrial settings, enabling machines to better interact with their physical environments.
Isaac 0.5 aims to allow vision-guided robots to navigate complex spaces like warehouses and factory floors, and to extract intelligence from recorded videos. The software is being released as an open-weight model, meaning its parameters and training materials are publicly accessible for inspection.
The company, which recently secured $21 million in funding led by Bessemer Venture Partners, was co-founded by Armen Aghajanyan and Akshat Shrivastava. They previously worked at Meta's Fundamental AI Research (FAIR) division. Perceptron positions its software as the future of industrial automated deployment, contrasting it with existing solutions that offer either generalist foundation models requiring significant cloud resources or narrow models that handle only perception or control.
According to the founders, Isaac 0.5 is a general-purpose tool designed for flexibility across various environments and situations, unlike specialized models. For instance, it can assist robots in tasks like reading package labels, understanding spatial arrangements, and planning pick-up sequences.
Models like Isaac 0.5 are trained on vast amounts of video data. Perceptron states its model was trained on a million hours of general video, supplemented by ego video (captured from a first-person perspective) and UMI video (recording repetitive human actions). The company has internally developed petabyte-scale datasets encompassing images, text, video, and robotic trajectories.
Perceptron believes its software has broad utility across industries including manufacturing, logistics, warehousing, security, mobility, and media and entertainment, positioning itself to lead the wave of industrial automation.