Infant Spontaneous Movement Noise Improves Exploration in Deep RL

This article presents a novel exploration mechanism in deep reinforcement learning (RL) that utilizes action noise inspired by infant spontaneous movements. The authors introduce a method that progressively increases the temporal auto-correlation of exploration noise during training, informed by the spectral properties of infants' end-effector velocities. Experimental results demonstrate that this infant-inspired noise enhances exploration efficiency and learning performance across various RL environments, suggesting that insights from human development can inform the design of more effective learning strategies for artificial agents.