Frontiers · World models and embodied AI
You can name the three live debates around what physical AI actually means in 2027 — understanding physics, embodiment-vs-simulation, and the robot-data bottleneck — and apply them to a frontier claim.
"Physical AI" is the phrase 2027 will be sold under. NVIDIA uses it, DeepMind uses it, and the humanoid companies all use it. The phrase carries a lot of weight that the underlying technology has not entirely earned yet.
This chapter is the unit's stress test. Three lessons, three questions:
Does the model understand physics? Three senses of "understand" are in play — pattern-matching plausibility (the model produces output that looks physically right), predictive accuracy (the model outputs match measurements), and causal reasoning (the model can answer counterfactuals). Most 2026 claims to "understanding physics" mean the first sense and are sometimes the second; few are the third.
Embodied versus simulated. What can only real-world embodiment teach? Contact mechanics at the millimeter scale, friction with unfamiliar materials, multi-step manipulation that accumulates small errors, social context in unstructured environments. Simulation can substitute for a lot but not all of this; the gap is where deployment surprises happen.
The robot-data bottleneck. Simulated data is unbounded but lossy. Real-world data is expensive but ground truth. Every physical-AI company has a bet on the right mix. Some bets pencil; some are theater.
Chapter contains 3 lessons.