Research map / Reinforcement learning / Visual and representation RL
Pretraining Representations for Data-Efficient Reinforcement 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 |
|---|---|---|---|
| Combine SPR, goal-conditioned RL, and inverse dynamics for offline pretraining | SGI · Atari 100k | median HNS: 0.343 -> 0.679 | Helped |
| Add goal-conditioned RL to inverse dynamics pretraining (ablation) | SGI-M · Atari 100k | median HNS: 0.411 -> 0.512 | Helped |
| Add SPR to inverse dynamics pretraining (ablation) | SGI-M · Atari 100k | median HNS: 0.411 -> 0.629 | Helped |
| Add inverse dynamics to SPR and goal-conditioned RL pretraining (ablation) | SGI-M · Atari 100k | median HNS: 0.029 -> 0.679 | Helped |
| Add SPR to goal-conditioned RL and inverse dynamics pretraining (ablation) | SGI-M · Atari 100k | median HNS: 0.512 -> 0.679 | Helped |
| Add goal-conditioned RL to SPR and inverse dynamics pretraining (ablation) | SGI-M · Atari 100k | median HNS: 0.629 -> 0.679 | Helped |
| Pretrain with Mixed rather than Random data | SGI · Atari 100k | median HNS: 0.326 -> 0.679 | Helped |
| Use the larger ResNet and 6M Mixed transitions rather than the standard SGI-M model | SGI · Atari 100k | median HNS: 0.679 -> 0.753 | Helped |
| Reduce learning rates for pretrained parameters during finetuning (ablation) | SGI-M · Atari 100k | median HNS: 0.429 -> 0.679 | Helped |
| Freeze the pretrained encoder during finetuning instead of using standard learning rates (ablation) | SGI-M · Atari 100k | median HNS: 0.429 -> 0.499 | Helped |
| Add SPR loss during finetuning (ablation) | SGI-M · Atari 100k | median HNS: 0.419 -> 0.679 | Helped |
| Use all SGI constituent losses instead of SPR alone during finetuning (ablation) | SGI-M · Atari 100k | median HNS: 0.679 -> 0.397 | Hurt |
Closest papers by meaning
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.91)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.90)
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration (ICML 2025 · similarity 0.89)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.89)
- The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
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.