Perceptron, a startup co-founded by former Meta researchers Armen Aghajanyan and Akshat Shrivastava, has released Isaac 0.5, a vision model aimed at enabling machines to interact more effectively with physical environments. The model is designed to assist vision-guided robots in navigating complex settings like warehouses and factory floors, while also extracting visual intelligence from recorded videos. The company emphasized that Isaac 0.5 is an open-weight model, allowing its parameters and training materials to be inspected by anyone. This approach contrasts with existing models that are often limited to specific tasks or require significant computational resources. The startup claims that its model is general-purpose, meaning it can adapt to different environments or situations rather than being confined to a single, repetitive task. According to the company, Isaac 0.5 is trained on a vast amount of video data, including general video, ego video, and UMI video, which captures human movements and interactions. The training data is described as petabyte-scale and spans multiple modalities, including images, text, video, and robotic trajectories. Perceptron aims to position itself as a leader in the automation wave by offering its software to a variety of industries, including manufacturing, logistics, and security. The company previously raised $16 million in 2024 and is currently in the process of closing an additional funding round.
Source: techcrunch
