A maintained glossary of physical AI terms — world models, VLA, sim-to-real, embodied AI — explained clearly and kept current as the field moves.
Physical AI
AI that perceives, reasons, and acts in the physical world through robots and autonomous machines.
Physical AI describes AI systems built to operate beyond text and images — perceiving their physical surroundings through sensors, reasoning about what to do, and acting on the world through motors, actuators, and robotic bodies. The term has been popularized industry-wide (notably by NVIDIA) to describe the combination of foundation models with robotics hardware: humanoids, industrial arms, autonomous vehicles, and drones. It's distinct from generative AI in that its output is physical motion, not a document or image, and it's typically judged on whether it works reliably in the messy, unstructured real world rather than on a benchmark.
Embodied AI
AI agents that learn and act through a body — physical or simulated — rather than through text alone.
Embodied AI is the research lineage behind physical AI: AI agents that perceive and act through a body, whether a real robot or a simulated one in a virtual environment. The concept traces back to embodied cognition research, which argues that intelligence is shaped by having a body that interacts with an environment, not just by processing abstract data. In practice, embodied AI and physical AI overlap heavily — physical AI tends to emphasize deployed, real-world hardware, while embodied AI is the broader academic term that also covers agents trained entirely in simulation before (or without) ever touching real hardware.
World Model
An AI's internal, learned simulation of how the world changes in response to actions — used to plan and predict before acting.
A world model is a learned internal representation of how an environment behaves — what happens if an agent takes a given action, how objects move, what's physically possible. Instead of learning purely from trial and error in the real world (slow and risky for a robot), an agent can "imagine" the outcome of an action inside its world model first, then plan or refine its behavior. World models are central to current physical AI research (examples include Genie-style interactive world simulators and driving-specific models like GAIA-1) because they let robots and autonomous vehicles train and plan in a compressed, predictive simulation of physics rather than only in the real world or a hand-built physics engine.
Sim-to-Real (Sim2Real)
Training a robot's AI in simulation, then transferring what it learned to work on real hardware.
Sim-to-real (often written Sim2Real) refers to both the technique and the core challenge of training a robot control policy inside a physics simulator — where trial and error is cheap, fast, and safe — and then deploying that trained model onto real robot hardware. The central difficulty is the "reality gap": simulated physics, sensor noise, and materials never perfectly match the real world, so a policy that works flawlessly in simulation can fail on real hardware. Techniques like domain randomization (varying simulated textures, lighting, and physics parameters during training) and increasingly realistic simulators are used to narrow that gap, making sim-to-real one of the practical bottlenecks physical AI companies work hardest to solve.