Building the humanoid brain, from perception to intelligence

15:35 - 15:50

Abstract

Humanoid robots will not become useful through bigger models alone: they need a nervous system, a tightly coupled stack of sensing, reflexes, and cognition that gives the brain reliable, low-latency access to the physical world. We argue that today’s vision-language-action (VLA) models fall short here, and make the case for a perception-language-action model that treats rich multimodal perception, including touch and proprioception, as a first-class input rather than an afterthought. Such models are only as good as the physical AI data they learn from, so we show how NEURA Robotics generates this data at scale in its robot gym, together with industrial and research partners. The talk closes with our roadmap from perception to intelligence and the open challenges where collaboration with the imec ecosystem can make the biggest difference.