Wheel-Hand-Leg-Arms Open Doors Playfully
spectrum.ieee.org
spectrum.ieee.org
Imagine that thing barrelling down a corridor at you to give you a hug, it's simultaneously adorable and terrifying.
Does anyone have different words than those used in the article which might unlock my comprehension?
You could give them no information and let them randomly move around until one happens to find the button.
But what if you just told them there was "something" interesting in a particular spot? You could skip hours and hours of training them just to find the button by just giving them a hint.
Think about the way a baby learns. They spot an object that catches their interest, and do whatever they can to get to it and inspect it. The difference is that the robot doesn't have an inherent mechanism for identifying interesting objects. You have to give it a hint as to what it should consider interesting, and it can figure the rest out from there.
The states are encoded in a feature space that ignores irrellivancies that don't effect and are out of the agent's control ( e.g. leaves blowing on trees )
Curiosity provides a method to learn this feature space from the raw pixel input via inverse dynamics.
The feature space is compact, relevant, sufficient, & stable which is ideal for RL.
This Curiosity Loss function entices the agent to explore states it has not been in and areas with complex dynamics.
These are states with high uncertainty which the agent reduces by becoming familiar with them.
This is a self-supervised method with no extrinsic human designed reward.
The agent aquires general knowledge about its enviroment and it is analagous to being curious.
Whooo lives in your nightmares and murders your friends?
Wheel-Hand-Leg-Arms!