A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot Control
Success Guided Sampling (SGS) is an adaptive sampler for Reinforcement Learning that focuses training on task configurations where the robot's policy is currently learning, rather than wasting experience on mastered or impossible tasks. By weighting sampling based on estimated success rates using a beta distribution, SGS concentrates RL training in moderately challenging areas, leading to better performance in large-scale simulated robot control tasks.