Armen Aghajanyan

Armen Aghajanyan was listed in the AI Engineer World’s Fair 2026 biography as co-founder and CEO of Perceptron AI and a former research scientist at FAIR and Microsoft. In his conference presentation, he described spending six years at FAIR working on scaling multimodal models, including early fusion, which brings different modalities together early in a model’s processing.

His presentation focused on Perceptron’s approach to embodied foundation models that combine perception, reasoning, and control. He described learning to predict useful future percepts rather than relying on sparse transcript supervision or treating every pixel as equally important. To reduce the computational burden of continuous visual input, he explained data sparse mixture of experts: a router selects or skips tokens across model layers, allocating computation toward informative regions and task-relevant objects.

Aghajanyan illustrated the team’s work with object detection that uses code, image tiling, zooming, contrast adjustments, and proposed bounding boxes; robotic video annotation that checks its own captions; and a book-sorting model that reads titles, identifies book types, and emits control tokens. He reported that joint training allowed ten times more video pretraining data to compensate for ten times less teleoperation data within the team’s tested compute range. He also acknowledged that temporal understanding remained insufficiently reliable for deployment and that robustness to changes in lighting and background remained limited.