How to train robots

If vision is still waiting for its GPT moment, robotics is a step further back, for a reason that has almost nothing to do with model size. The wall in front of physical AI is evaluation rather than intelligence. You can train a policy overnight, but you cannot test it overnight, because the test happens in the physical world and the physical world runs at one times speed. You cannot speed up reality.

In digital AI, iteration is cheap: generate, score, repeat, a million times before lunch. In robotics, every iteration costs a real robot doing a real thing in real time, and the scoring is slow, expensive, and hard to standardise. The teams that win will industrialise evaluation, not just train the biggest model.

Two iteration loops, orders of magnitude apart Digital AI Generate a policy, score it against data, repeat. The environment is free and instant. Each mark is one evaluation run. ~10⁶ / day Iterations are effectively unlimited. Physical AI Every trial is one robot doing one real task, scored slowly and by hand. Same window. One run. 1× real time You cannot speed up reality.

The tyranny of the extra nine

The arithmetic is unforgiving. A task with a hundred sequential steps, where each step succeeds ninety-nine percent of the time, succeeds as a whole about thirty-seven percent of the time. Ninety-nine percent per step sounds excellent; it produces a machine that fails two times in three.

Reaching ninety-nine percent success across the whole task requires something closer to four nines, 99.99 percent, at every single step. Each additional nine is a brutal, expensive climb, and each one marks the difference between a demo that goes viral and a product someone will actually deploy. Most robotics you see online lives at one or two nines. The gap between that and a shippable system is years of reliability engineering, not a better prompt.

The tyranny of the extra nine 37% 90% 99% 99% / step 99.9% / step 99.99% / step 0% 25% 50% 75% 100% Chance the full 100-step task succeeds

Data or architecture

There is a real fight about where the unlock comes from, and it splits into two camps:

I find the strongest evidence for the architecture side comes from biology. A mosquito navigates, hunts, and avoids your hand on roughly a tenth of a watt, with no pretraining run and no data center behind it. Whatever it is doing, it is not brute-forced scale, which suggests that a far more efficient approach to physical intelligence is possible and that pouring more data into the current designs may be climbing the wrong hill.

Where robotics actually is

Robotics is at its GPT-2 moment: qualitatively different from a year ago, good enough to excite and to convince serious people it is close, yet still a full generation short of the breakthrough that turns a research curiosity into deployable infrastructure. GPT-2 was thrilling and commercially useless, and that distance between capability that impresses in a clip and capability that survives contact with a paying customer is roughly where the robots sit today.

Take the capital already committed as the measure of how early this is. Figure raised over $1 billion at a $39 billion post-money valuation, Apptronik more than $935 million, Agility carries a $2.5 billion valuation against $300 million or so of multi-year orders, and Unitree is targeting a listing near $6.2 billion on a China cost curve none of the Western platforms can match. Waymo is the one that has crossed over, running more than 500,000 paid rides a week, and it took roughly fifteen years and Alphabet's balance sheet to get there in a single, unusually structured task. The last point is the one investors should sit with, because it echoes what is true in vision. The advantage that compounds is access to the real-world interaction data rather than the model itself, and that access turns out to be a commercial and relationship problem rather than a technical one. The two-sided market being built on the data side of that debate is a bet on exactly this, that whoever can strike the deals to collect real data at scale, in real environments, ends up with the asset nobody can clone. The pipeline into reality is the moat, not the training recipe.

The views cited belong to two founders building in robotics. The framing and any mistakes are mine.