Perceptron AI

Perceptron AI builds foundation models for understanding and acting in the physical world. Its Mk1.5 model processes text, images, video and audio, letting developers track objects, transcribe recordings, extract structured information and connect models to application tools. Available through its platform and Python SDK, it supports agents on robots, drones, smart glasses and phones. Its Isaac model family addresses robot control: teams can fine-tune models on demonstrations or incorporate their visual outputs into planners and controllers.

Launched by Armen Aghajanyan and Akshat Shrivastava, Perceptron combines multimodal perception with embodied reasoning and action. Its Isaac research trains a shared backbone on video understanding and robot control, including predictions of future task-relevant states such as object relationships or contact changes. This connects learning from video to generating actions. Its sparse architecture also varies the number of active experts according to task complexity, adjusting computation within the model.

Developers can use hosted APIs, while Isaac’s released weights and training code support customization for robotics. For workflows that repeatedly examine the same media, Multilook processes a shared image or video context once for several questions in one request. Applications include annotating robot demonstrations, inspecting industrial procedures and checking multiple attributes of a retail shelf.