Too many moves, too few labels
World models · 2026 · Part 3 of a series that starts with The drone and the world model
Search worked for board games because the moves were few and the rules were known. This page counts what a car or a robot faces instead, and asks where the data to learn it could come from.
World models · part 3
Too Many Moves, Too Few Labels
Search beat chess and Go because there were few moves and perfect rules. A car or a robot has neither. How many moves does a robot really have, and where would the data to learn them come from?
This page counts the action spaces of cars and robots, separates the two kinds of data, and sets out the two ways to turn data into behaviour. Wherever a shortcut version is commonly said or written, it's marked COMMON SLIP and the correct form is shown.
1 · Counting the moves
A car has more moves than Go. A robot has far more.
Part 2 showed that tree search needs a short list of moves. Go has 362. To compare, chop a car's controls into steps: so many steering angles, so many brake levels, so many throttle levels. Every combination is one “move”. Do the same for a robot arm, one joint at a time.
Lab 1 · Chop the controls into steps
every number below is computed from the slidersCar
Robot arms
Number of moves, log scale (each tick is ×10)
Common slip: “2 degrees of freedom per joint”
An industrial arm's joints each turn about one axis, so a “6-axis arm” has 6 degrees of freedom in total, not 6 per segment pair. A two-armed robot with 7 joints per arm and a simple one-motion gripper on each has (7 + 1) × 2 = 16. With 10 steps per joint that is 1016 moves. The famous total survives; the “2 + 2 + 2 + 2 per arm” route to it doesn't.
The car count usually quoted is 365 × 10 × 10 = 36,500, about 100 times Go. A circle has 360 degrees, a real steering wheel turns about 1.5 turns each way, and nobody presses both pedals hard at once, so the honest count moves around a lot. Every version is far beyond what a tree search can visit, and the true controls are continuous anyway.
2 · Two kinds of data
Video shows what happened, not what anyone did
Passive video (dashcams, phone clips, cooking shows) is a list of states and nothing else:
Action-labelled data records the controls too, and sometimes a reward:
Video is nearly free and nearly endless. Labels are expensive, except in one lucky case: a car fleet. Every customer car already logs its camera images and the driver's steering and pedals, so each mile driven is labelled data. Robots have no such fleet. Their labelled data comes from teleoperation, a person steering the robot by remote control, one slow demonstration at a time.
Labels also don't transfer cleanly between bodies. “Brake at 30%” stops a heavy SUV differently from a light hatchback, and the same joint angles put two different robot hands in different places. That's the cross-embodiment gap: the action means something different on a different body.
3 · Two ways to act
Map pictures to moves, or imagine first
Common slip: “a VLA is model-free RL”
Model-free, yes: a VLA has no world model. But most VLAs are not trained by reinforcement learning at all. They are trained by imitation (behaviour cloning): copy the action a human teleoperator took in each recorded state. That is supervised learning. Reinforcement learning, model-free or not, learns from rewards earned by trying things. For how all these pieces fit together, see World models vs RL.
4 · Next
Video is plentiful. Can it stand in for labels?
A world model needs to know what actions do, which takes labelled data. But most of what it must learn, how the world moves on its own, is in ordinary video. The next page shows how far watching gets you, and how few labels you need after it.
Based on minutes 40–45 of the YC Decoded video “World Models: An intuitive introduction” (July 2026). The equations were checked against the video's whiteboard and corrected where needed.
Sources. Degrees of freedom · A two-armed 7-axis robot (YuMi) · RT-2: an early vision-language-action model (2023) · π0: a VLA trained by imitation (2024) · Open X-Embodiment: data across 22 robots (2023)