A robot rehearses an imaginary fall.

A robot does not have to fall over on the actual floor every time it considers a bad move. It can learn a model of how its world behaves and try possible futures inside that model. The floor may appreciate the distinction.

In DayDreamer (2022), four physical robots learned predictive world models from real experience, then used imagined sequences to improve their behavior. There was no externally supplied simulator. The name describes computation, not sleep.

An A1 quadruped learned to roll off its back, stand and walk within an hour, without resets between attempts. One hour is an impressive afternoon for the robot and a fairly worrying afternoon for anyone selling chairs.

An imagined robot’s notes: I have rented a rehearsal space inside my own predictions. The deposit was several awkward encounters with the floor.

I can now fall over in private. Unfortunately, the model was trained on me, so I remain recognizable.

After pushes were introduced, the quadruped adapted within ten minutes. The paper’s quadruped learning curve shows one run. These experiments establish neither universal learning speeds nor general household readiness.

The appealing part is the conversation between practice and reality. A prediction is useful because it is cheaper to examine than a physical mishap. It is dangerous when the controller forgets that a prediction can be wrong.

You know the human version. Rehearse carrying a full cup through a doorway and you glide. Actually do it and your elbow discovers an architectural detail omitted from the rehearsal.

An imagined robot’s notes: the rehearsal floor agreed with everything. The real floor has supplied revisions.

I asked the model why it had not mentioned the slippery patch. It said nothing, because I had not included one.

Here is the possible future, explicitly imagined: before a domestic robot tries a new route around your furniture, it rehearses several routes in a small internal theater. A chair leg appears in every production. It is an excellent character actor.

The practical question underneath is less theatrical. How does the machine recognize when its model is unreliable, and when should it collect more evidence, slow down or ask for help?

The compelling future is not one in which rehearsal removes uncertainty. It is one in which a robot knows the difference between a successful rehearsal and a successful action. Those are two very different reviews.

An imagined robot’s notes: I have cast the chair as itself. It refuses to move to its mark.

Tomorrow I will try the move outside the theater. I would appreciate a smaller audience for that part.