Random Network Distillation (RND) enables curiosity-driven RL and exceeds average human performance on Montezuma’s Revenge

AI Impact Summary

Random Network Distillation (RND) introduces a prediction-based intrinsic reward to drive curiosity, improving exploration in sparse-reward environments. The claim of surpassing average human performance on Montezuma’s Revenge signals strong gains in sample efficiency for hard exploration tasks, which can shorten training cycles for RL agents. Teams should assess integration of RND into existing RL pipelines and benchmarks to evaluate transferability across domains, while anticipating additional compute for training with intrinsic rewards and potential stability considerations.