Physical AI, Chapter 10: The Road to Embodied AGI

This is the final post in a series walking through my book Physical AI: Engineering Embodied Intelligence for the Real World. In the previous chapter, we looked at what has actually shipped. This one closes the book by asking where the field goes next.

Three trajectories that ran in parallel for a decade have converged. Foundation-model capability in software: transformers, then scaled language models, then multi-modal vision-language models, then their application to robotics through RT-1, RT-2, Open X-Embodiment, Pi-0, Helix and GR00T. Mature supply chains in hardware: brushless motors as a commodity since the 2010s, and harmonic drives moving from a handful of specialty manufacturers to broad availability, taking the per-joint cost down with them. And capital at a scale the field has not seen — Figure at $675 million in early 2024 and a valuation north of $30 billion since, Physical Intelligence at $400 million, tens of billions in aggregate across venture, corporate and state sources.

None of the three is individually new. What is new is that they are now intersecting, and that they are mutually dependent in a way that makes the whole vulnerable to a setback in any one of them. A failure of the VLA scaling hypothesis, a supply-chain reversal in actuator cost, or a capital-market correction would each independently slow the convergence.

Reading the forecasts honestly

Goldman Sachs put the 2035 humanoid TAM at $6 billion in 2022, then revised it in January 2024 to $38 billion base case and $154 billion bull case — mostly on a lowered bill-of-materials assumption, from $250,000 per unit to around $150,000. The chapter takes the forecasts seriously enough to examine what they actually assume: a unit-cost curve dropping toward $20,000–$50,000 by the early 2030s, an adoption curve that starts industrial and progresses to consumer, labour displacement as the dominant demand driver, and geopolitical stability sufficient to keep the East Asian component supply chain intact. Each assumption is plausible; none is guaranteed; and the bull case requires the consumer market to materialise, which is the most speculative claim in the set. Read them as conditional scenarios rather than predictions.

The geopolitical picture cuts across the usual framings. The US leads in foundation models and capital; China leads in hardware manufacturing, has the deeper humanoid portfolio, and has made the segment a national priority — Unitree’s G1 launched below $20,000, alongside Fourier, XPENG, UBTECH and Agibot. A US humanoid program sources Chinese components; a Chinese program uses US-developed foundation-model techniques. The interdependence is the strategic fact, and both sides are exposed to it.

What is actually unsolved

Long-horizon planning, where today’s VLAs operate over seconds and the eventual deployments need hours — a problem that decomposes into credit assignment across long delays, memory architecture, and recovery from sub-task failure without restarting. Multi-robot coordination at time scales tighter than a central task allocator supports. Deployment in unstructured environments and the long tail behind it. Whether scaling laws hold for robotic policies at all: the evidence through 2026 is suggestive, but there are not enough data points to fit a Chinchilla-style curve with any confidence, and the answer will shape the next decade. Energy and cost, where a humanoid draws 400–800 watts for four to six hours and costs $10–30 an hour fully loaded. And the two that engineers habitually underrate: the workforce and social-acceptance dimension, and the integration of the robot with the EHR, MES or WMS it has to live alongside — which is where most deployments actually get hard.

The chapter closes with six entry points for practitioners who want to contribute without joining a capital-intensive program: the open-source model ecosystem, the simulators, dataset collection, hardware, application engineering, and the standards and policy work that will govern all of it. Each has a different time horizon, and the honest advice is to match the horizon to your circumstances.

That closes the series and the book. We went from the omelette a robot cannot make, through the history that produced the current architectures, the stack, the policy, the data bottleneck, the simulators, the edge hardware, the safety apparatus, and the industries where this is already working — to the questions nobody has answered yet. The rest is the part no book can hand you: the judgment to apply it to a system you actually own. Thank you for reading along.