Perceptron Unveils Isaac 0.5, Open-Weight Visual Model

Founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava (former FAIR), Perceptron introduces Isaac 0.5, a vision model released with open weights. According to the company, the training accumulated one million hours of videos and relies on in-house datasets at a petabyte scale. Funded with $21 million, the startup aims for industrial deployments and is preparing for commercialization with multiple suppliers.

One Million Claimed Hours and Internal Datasets

Perceptron claims to have fed Isaac 0.5 with one million hours of so-called general videos to learn to recognize environments, visuals, and scenarios. At the same time, the company states that it will not disclose the precise origin of this training data. Akshat Shrivastava specifies that internal datasets, at the petabyte scale, were built by combining images, text, video, and robotic trajectories. Models of this type learn by ingesting vast video corpuses. In this case, the company also mentions that it relied on egocentric videos, captured from a first-person perspective, as well as UMI videos used to teach movements by recording repetitive human actions.

Announced Navigation and Visual Analysis Capabilities

According to Perceptron, the software can assist vision-guided robots in navigating warehouses or along production lines, and help companies extract information from the videos produced by these robots. To illustrate a concrete case, Akshat Shrivastava cites package sorting: a robot should first read a label, analyze the space to locate boxes, decide which one to pick up, and then, in the case of a series, plan the order of retrieval. The company presents Isaac 0.5 as a tool designed to support each step of these sequences. It notes that software already exists to perform most of these tasks, while arguing that few have been designed to do so flexibly. Perceptron further argues that in physical robotics, the current choice is between generalist models that are GPU cloud-hungry and narrow models limited to perception or control; its co-founders claim, on the contrary, to aim for a versatile model capable of adapting to the situation.

Announced Commercialization and Targeted Sectors

The startup claims to be able to make its software available to various suppliers and believes that its intelligence component is likely to integrate into many sectors. Targeted sectors include industrial production, logistics and storage, security, mobility solutions, as well as the media and entertainment industries. The founders see their software as the next step for industrial automation, continuing their approach to spreading across multiple markets.

Launch, Funding, and Team Behind Isaac 0.5

Perceptron launched Isaac 0.5 this week and released it with open weights, with parameters and training materials available for review. The company recently raised $21 million in a round led by Bessemer Venture Partners. Founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former members of Meta's FAIR lab, Perceptron develops vision models dedicated to interacting with the physical world. According to the company, Isaac 0.5 is designed to enable machines to perceive, reason, and act in industrial environments. "Nothing like this really exists," says Armen Aghajanyan, who expresses enthusiasm.