Research map / Reinforcement learning / Visual and representation RL
Large-Scale Study of Curiosity-Driven Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add intrinsic curiosity reward to extrinsic reward | PPO · Sparse-reward Atari games | extrinsic reward: Gravitar: 999.3 \pm 220.7 -> 1165.1 \pm 53.6; Freeway: 33.3 \pm 0.6 -> 32.8 \pm 0.3; Venture: 0 \pm 0 -> 416 \pm 416; Pr | Mixed |
| train with curiosity reward and no extrinsic reward | Curiosity-driven agents · Atari games | extrinsic reward versus random agent: IDF: 75%; RF: 70% | Helped |
| use inverse-dynamics features instead of random features (ablation) | Curiosity-driven agents · Atari games | extrinsic reward comparison: 55% | Mixed |
| replace feature-space dynamics prediction with raw-pixel prediction (ablation) | Dynamics-based curiosity · Atari games and other evaluated environments | performance | Hurt |
| use VAE features instead of random or inverse-dynamics features (ablation) | Dynamics-based curiosity · Atari games and Super Mario Bros. | performance | Mixed |
| increase parallel environment threads from 128 to 2048 (ablation) | RF curiosity · Super Mario Bros. | exploration and game performance: 128 to 2048; 11 different levels | Helped |
| pre-train curiosity on Level 1-1, then transfer to Level 1-2 (ablation) | RF and IDF curiosity · Super Mario Bros. Level 1-1 to Level 1-2 | transfer performance | Helped |
| pre-train on Level 1-1 and transfer across a day-to-night color shift (ablation) | RF curiosity · Super Mario Bros. Level 1-1 to Level 1-3 | transfer performance | Hurt |
| use learned inverse-dynamics features rather than random features (ablation) | Curiosity-driven agent · Super Mario Bros. novel levels | generalization | Helped |
| add a noisy TV with action-dependent random channels (ablation) | IDF and RF curiosity · Unity maze with terminal reward | learning speed | Hurt |
| add curiosity reward to terminal extrinsic reward | PPO in the Unity maze · Unity maze with terminal reward | goal attainment | Helped |
Closest papers by meaning
- Curiosity in Hindsight: Intrinsic Exploration in Stochastic Environments (ICML 2023 · similarity 0.91)
- Curiosity-driven Exploration by Self-supervised Prediction (ICML 2017 · similarity 0.91)
- Wonder Wins Ways: Curiosity-Driven Exploration through Multi-Agent Contextual Calibration (NeurIPS 2025 · similarity 0.89)
- DISCOVER: Automated Curricula for Sparse-Reward Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Can In-Context Learning Support Intrinsic Curiosity? (NeurIPS 2026 · similarity 0.87)
- MaxInfoRL: Boosting exploration in reinforcement learning through information gain maximization (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.