Intrinsically Motivated Learning
Regularity as Intrinsic Reward for Free Play
RaIR
We propose Regularity as Intrinsic Reward (RaIR) for model-based reinforcement learning, inspired by child development, to guide exploration towards regular situations otherwise not easily discovered.
Inspired by child development, we postulate that the search for structure partially guides exploration. RaIR is a measure of regularity of a scene and is computed using Shannon Entropy of a suitable scene representation.
By combining RaIR with epistemic uncertainty, our approach enables autonomous construction of regular structures during free play, significantly boosting zero-shot performance on assembly tasks.
Publications
Regularity as Intrinsic Reward for Free Play
Advances in Neural Information Processing Systems 37
·
2023