Perceptron, a start-up company founded by former Meta AI researchers, released a new generation of visual models, Isaac 0.5, with the goal of bringing AI further from digital scenes into the operational environment of entities such as storage, factories, etc. According to the company, the model can help robots to sense, judge and implement in complex industrial scenarios.

To store and plant robots.

Perceptron was founded in November 2024 and co-founded by Armen Aghajanyan and Akshat Shrivastava, who served in FAIR, Meta Basic AI Research Department. Both wanted to create a more universal visual AI system for industrial automation deployment.

According to the company, Isaac 0.5 can be used to provide visual guidance for robots to move in warehouses or workshops and to process live video messages from which visual data can be extracted for operational analysis.

Unlike a model for training only for a single duplicate, Perceptron tries to make Isaac 0.5 a generic system. According to the company, the model could be adapted to different environments rather than only for a fixed process.

Training data to cover video and robot tracks

Perceptron states that Isaac 0.5 uses a generic video of about 1 million hours for training to enhance model recognition of scenes, objects and operational situations.

  • First perspective video, the human operation video.
  • UMI video data for learning to repeat actions
  • Image, text, video and robotic trajectory data

The company did not disclose specific sources of such training data, but the co-founder stated that the team had built an internal PB-level data set to support model training.

Financing of $21 million recently completed

Perceptron recently completed $21 million in financing, with Bessemer Venture Partners leading the process. The company is preparing to move the software to a larger number of industry clients in the hope of connecting its intellectual layer to a wider automated system.

  • Manufacturing
  • Logistics and storage
  • Security, mobile travel and media entertainment

Perceptron considers that the options that are common in the current entity AI field are either based on a cost-effective generic base model or can only use a narrow single-point model. The objective is to provide a programme that is sensitive and controlled and that can be deployed flexibly in different contexts.