From VLM/VLA's to Embodied Agents — Armen Aghajanyan, Perceptron AI

Why watch

Armen Aghajanyan presents Perceptron's approach to embodied agents, addressing two core problems: wasted visual tokens in video training (98% lack ground truth) and context bloat from continuous cameras. The solution combines a perceptive objective that learns which percepts matter with a data sparse mixture of experts router that selectively processes tokens per layer. Their model trained on 1PB of multimodal data (text, images, video, trajectories) beats frontier lab embodied reasoning models at lower cost and demonstrates a new scaling law where 10x video pretraining substitutes for 10x less expensive teleop data.