This page answers questions a reader may still have about the world model in this book. Each answer is short. It links to the section that explains it in full.
Contents#
- What does the model predict here that it did not predict in the pushing book?
- Why did its training data cost 22 arm hours?
- Why does it not push twice from one look?
- It separates the most tables, so why is its share not the best?
- Why was a change kept that broke more glasses on the held-out queue?
- When it breaks a glass, which glass is it?
1. What does the model predict here that it did not predict in the pushing book?#
It predicts the reach. That is how far the hand goes in before it touches the glass. A path cannot feel for the glass, so the reach must be known first. The reference finds it by poking, which costs arm time. Why a prediction is worth more here explains the trade.
2. Why did its training data cost 22 arm hours?#
Without the overhead camera, the arm must walk round the table to see it. It walks the ring before it knows where the glasses are. It walks again when a pushed glass is hidden. So each collecting table costs about 28 arm seconds for four or five pushes. What the data cost the arm gives each round.
3. Why does it not push twice from one look?#
It was measured, and it saved nothing. The look after a push is usually free, because the jaw already stands behind the glass. Pushing again from the model's belief cost about 10% more arm time for the same tables. Why it still looks after every push has the table.
4. It separates the most tables, so why is its share not the best?#
A table with a broken glass earns nothing, and it breaks some. It also reaches many tables and leaves some of them crowded. On the 26 tables both solutions reached, it ties the reference. Why the share is lower splits the gap between those two causes.
5. Why was a change kept that broke more glasses on the held-out queue?#
The change was twice the pushes, with a heavier charge for a missed topple. It was chosen below the held-out line, where it broke fewer glasses. The project does not undo a choice because the held-out queue disagrees. The second repair gives both sets of numbers.
6. When it breaks a glass, which glass is it?#
It is always the glass being pushed. Over 300 tables, 30 ended with a glass over, and on none of them did a neighbour fall. So the fault is a glass tipping instead of sliding. Which glass goes over says what that means for the next step.
← What the repairs bought · A foundation model as it downloads — what is different →