Foundation models are bringing general intelligence to physical robots. Vision-Language-Action (VLA) models combine visual perception, language understanding, and motor control in a single model. RT-2 (Google DeepMind): fine-tunes large VLM (PaLI-X) to output robot actions as text tokens - "move arm right 5cm" becomes action tokens. Trained on web data + robot demonstrations, generalizes to novel objects and scenarios not in training data. OpenVLA: open-source 7B VLA trained on Open X-Embodiment dataset covering 22 robot types, 97K episodes. Performance: 28.9% absolute improvement over diffusion-based policies on generalization tasks. Physical Intelligence (pi): commercial company building general-purpose robot foundation models, raised $400M. Their pi0 model achieves unprecedented cross-task generalization. Key challenges: simulation-to-reality gap for data generation, long-horizon task planning, manipulation of deformable objects, sample efficiency. Data collection: teleoperation (human demonstrates, robot learns), simulation with domain randomization, real-world deployment data. Current state: robots can now perform complex household tasks following natural language commands, though still brittle in unstructured environments. 2025 expectation: reliable performance on structured manufacturing and warehouse tasks.