Experiment · Physical Intelligence
What we are learning by taking operations into the physical world
The difference from a digital operation is not the difficulty of the control, but the fact that the state has to be estimated.
This is an exploration from the Laboratory, over controlled scenes and with a deliberately small scope. What follows are observations, not conclusions.
We expected the hard part to be the control. It has been perception, and not because of accuracy: because of its nature. A sensor does not deliver the state of the environment, it delivers a noisy measurement from which the state is estimated. The whole design changes once you accept that the system works over a hypothesis and not over a reading.
The second observation is that time stops being an implementation detail. An estimate expires while you are deciding what to do with it. In a digital system a slow query is an annoyance; here a late decision is applied to a world that has already become another one.
The third has surprised us less, but it weighs more: almost nothing is reversible. Undoing a record is cheap; undoing a movement rarely is. That puts the question of safety before the question of performance, and forces you to declare a default behaviour for each failure mode before writing the main logic.
We still have no satisfactory way of estimating when perception is insufficient. We discover it from the result, which is late.