Regularity as Intrinsic Reward for Free Play
Cansu Sancaktar, Justus Piater and Georg Martius
RaIR: Regularity as Intrinsic Reward
RaIR: Regularity as Intrinsic Reward
Optimizing for RaIR with Ground Truth (GT) models gives us regular and symmetric patterns/constellations!
RaIR + CEE-US: We inject RaIR into free play together with ensemble disagreement.
RaIR + CEE-US: We inject RaIR into free play together with ensemble disagreement.
Free Play with RaIR + CEE-US in Construction Environment
Free Play with RaIR + CEE-US in Construction Environment
Some snapshots from free play
Iteration 271
Iteration 285
Iteration 292
Iteration 295
Patterns generated by optimizing for RaIR including color as a symbol in ShapeGridworld
Patterns generated by optimizing for RaIR including color as a symbol in ShapeGridworld
Poses generated by optimizing for RaIR in locomotion environments with GT models
Poses generated by optimizing for RaIR in locomotion environments with GT models
Quadruped
Walker
Free Play in Quadruped with RaIR + CEE-US
Free Play in Quadruped with RaIR + CEE-US
Free play iteration 110
Free play iteration 150
Free play iteration 165
Free play iteration 172
Examples for Zero-shot Downstream Task Performance
Examples for Zero-shot Downstream Task Performance
(Videos are RaIR + CEE-US solving the tasks)
Balance Front
Stand Rotated
Attack
Balance Back
Free Play in Custom Construction (diverse shapes) with RaIR + CEE-US
Free Play in Custom Construction (diverse shapes) with RaIR + CEE-US
Free play iteration 150
Free play iteration 175
Free play iteration 180
Free play iteration 190
Examples for Zero-shot Downstream Task Performance
Examples for Zero-shot Downstream Task Performance
(Videos are RaIR + CEE-US solving the tasks)
Stack Cube + Ball
Stack Cube + Flat Block
