Former Meta researchers launch Perceptron to automate industrial robotics

The company recently secured $21 million in funding led by Bessemer Venture Partners to scale its vision-guided technology. Unlike existing industrial programs limited to repetitive tasks, Isaac 0.5 is built to handle multiple steps of a workflow—from spatial analysis and label reading to complex picking sequences—without requiring dedicated cloud GPU clusters for every individual robot instance. By releasing the software as an open-weight model, the founders invite external scrutiny of its parameters and training architecture.

To achieve this level of autonomy, the model was trained on a massive dataset comprising a million hours of general video, alongside specialized 'ego' and 'UMI' footage. These inputs allow the system to mimic human perspective and repetitive physical actions. While the specific data sources remain proprietary, the team claims to have constructed petabyte-scale datasets spanning images, text, and robotic trajectories. Perceptron intends to license this intelligence layer to vendors across manufacturing, logistics, security, and entertainment, aiming to replace narrow automation with a more adaptable, general-purpose layer for the physical world.