From Rolling Over to Walking: Enabling Humanoid Robots to Develop Complex Motor Skills

We present a novel method for enabling humanoid robots to learn a wide range of motor skills through reinforcement learning. Our approach introduces an achievement-triggered multi-path reward function that draws on principles from developmental robotics, enabling the robot to learn a majority of the gross motor skills typically acquired by human babies in a single training stage. Our experiments in a simulation environment demonstrate higher success rates and faster learning compared to traditional reinforcement learning approaches. Our method leverages self-discovery and exploration, mimicking human infant learning, and has the potential to significantly advance the state-of-the-art in humanoid robot motor skill learning.