A seminal 2018 study by OpenAI researchers Yura Burda, Harri Edwards, Deepak Pathak, Amos Storkey, Trevor Darrell, and Alexei A. Efros conducted the first large-scale investigation into purely curiosity-driven reinforcement learning. By utilizing prediction error as an intrinsic reward signal, the team explored agent performance across 54 benchmark environments, including the Atari game suite, without any external rewards.

The findings revealed that curiosity-driven learning achieved surprisingly strong performance, closely aligning with hand-designed extrinsic rewards in many games. The research also analyzed the impact of different feature spaces, determining that while random features suffice for many RL benchmarks, learned features offer superior generalization, particularly to novel levels in games like Super Mario Bros. Additionally, the study identified limitations of prediction-based rewards in stochastic environments.

This work is critical for the AI industry as it addresses the scalability challenge of manual reward engineering, paving the way for more autonomous and adaptable AI agents capable of learning complex behaviors in unstructured real-world scenarios.