Physical AI: How Robots Are Moving Beyond Scripted Moves

Physical AI refers to the intelligent software that lets robots perceive, understand, and act in the physical world, distinct from the humanoid form factor. While humanoid robots often perform impressive scripted demos, they lack the ability to adapt to changing conditions. The industry is shifting from training a single massive end-to-end model to an agent-based approach, where multiple specialized models handle different modalities (vision, language, action) and can be updated independently. Learning remains a bottleneck: there are no large-scale datasets like those used for LLMs, and methods such as learning from video still require supplementary annotations. Vision is the most mature sensory channel, with Vision-Language-Action (VLA) models leveraging pre-trained vision encoders and tokenizing images like text. Language acts as a universal interface, allowing high-level commands to be broken down by high-level planners. Tactile sensing is deemed critical for tasks requiring precise manipulation, as vision is insufficient once an object is gripped; research explores artificial skin, but current robots use force sensors. Proprioception—knowing the robot's own body state—is another essential channel, using encoders, IMUs, and torque sensors. Achieving human-like dexterity and adaptability remains a distant goal, but agentic architectures and incremental improvements are seen as the practical path forward.