Robotic foundation models are large pretrained systems adapted to produce robot actions from sensory input and language goals. The bet is that internet-scale vision and language can transfer into control—so a robot can see a scene, understand an instruction, and act without being hand-engineered for every task.
Inside embodied intelligence, this chapter focuses on model families, datasets, and fine-tuning recipes rather than low-level control theory. The important subproblems are action tokenization, multimodal alignment, safety constraints, and sim-to-real transfer. The important methods are vision-language-action models, behavior cloning at scale, and co-training with web data.
What is still unclear is how much generality is real versus benchmark-specific, and how to evaluate long-horizon competence once demos look good. On the field map, this feeds the Learning section of Embodied Intelligence and connects to paper breakdowns like RT-2.