Do Generalist Robots Need Specialist Models?

Large Vision-Language Models (VLMs) have demonstrated impressive generalization in the digital realm, but translating this into reliable robot manipulation and navigation remains a fundamental challenge. This talk explores a hybrid path forward: augmenting generalist "brains" with specialist "nervous systems." I will first present two foundation model efforts: SeeDo, which leverages VLMs to interpret long-horizon human videos and generate executable task plans, and INT-ACT, an evaluation suite that diagnoses a critical intention-to-execution gap in current Vision-Language-Action (VLA) systems. This gap reveals a key generalization boundary: robust task understanding does not guarantee robust physical control. To bridge this divide, I will introduce specialist models that provide two missing ingredients: fine-grained physical understanding and acquiring data for learning at scale. EgoPAT3Dv2 grounds robot action by learning 3D human intention forecasting from real-world egocentric videos. To address the data-scaling challenge, RAP employs a real-to-sim-to-real paradigm, while CityWalker explores web-scale video to learn robust, specialized skills. I will conclude by drawing analogies from the only known generalist agents — ourselves — to offer my answer to the question posed in the title.