Visual AI: From Research to Practice on the Factory Floor
Artificial intelligence is everywhere, but it still mainly lives in the digital world. However, that is changing. More and more, startups are trying to bring AI into the real world. And that’s where Perceptron comes in, a company founded by two former Meta scientists, aiming to transform the interaction of machines with the physical environment.
Founded in November 2024, Perceptron develops cutting-edge vision models to help machines better interact with their physical environments. This week, they launched Isaac 0.5, a model designed to give machines the ability to "perceive, reason, and act" in industrial settings. Imagine vision-guided robots navigating through warehouses or complex factory floors. That’s what we’re talking about.
The interesting detail is that Isaac 0.5 is an open-source model, allowing anyone to inspect its parameters and training materials. This is a bold move and opens doors for collaborative innovations.
The Founders' Vision
Armen Aghajanyan and Akshat Shrivastava, the co-founders of Perceptron, came from Meta's AI research team, FAIR. They believe their software is the future of industrial automation. According to them, today’s physical AI forces a false choice: either you have generalist models that require multiple dedicated GPUs in the cloud, or limited models that handle perception or control, but never both.
What makes Isaac 0.5 different is its flexibility. It is not made for a specific repetitive task. Instead, it adapts to the environment or situation it finds itself in. Imagine a robot organizing boxes. It needs to read labels, analyze space, and decide the order of actions. Perceptron’s software helps robots navigate each step of this process.
The Magic of Data
But where does the "magic" of this algorithm come from? Models like Isaac 0.5 learn operational skills by consuming vast amounts of video data. Perceptron fed its new model with a million hours of general videos to teach the algorithm to identify specific scenarios and visuals. They also used "ego" and "UMI" videos to teach movements, capturing repetitive human actions.
Although Perceptron does not disclose the exact sources of its training data, Shrivastava mentions that the company built internal petabyte-scale datasets encompassing images, text, video, and robotic trajectories.
The New Era of Automation
The utility of software that helps robots operate in warehouses is immense. Perceptron believes it is well-positioned to lead this wave of automation. They are ready to market their software across various sectors, potentially integrating their intelligence layer into industries such as manufacturing, logistics, security, mobility, media, and entertainment.
Nothing like this really exists out there, says Aghajanyan. And the excitement is palpable.
The startup has already raised $16 million from investors like Bessemer Venture Partners and SmartGateVC in 2024. And, it seems, they are closing another funding round.
AI is coming out of screens into the real world, and Perceptron is at the forefront of this transformation. The factory floor will never be the same.
Comments (0)
Comments are moderated and if they violate our Terms and Conditions of use, the comment will be deleted. Persistence in violation will result in a ban of your account.