Research map / Reinforcement learning / Visual and representation RL
Search on the Replay Buffer: Bridging Planning and Reinforcement Learning
Ben Eysenbach, Ruslan Salakhutdinov, Sergey Levine
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 graph search over replay-buffer observations as intermediate waypoints | Goal-conditioned RL · 2D navigation with goals over 100 steps away | success rate | Improved |
| Add graph search over replay-buffer observations as intermediate waypoints | Goal-conditioned C51 · Visual Navigation | success rate: around 90% | Improved |
| Add graph search over replay-buffer observations as intermediate waypoints | Goal-conditioned RL · SUNCG held-out houses, goals 10 steps away | success rate: almost 80%; less than 20% | Improved |
| Use SoRB's learned distance predictions instead of SPTM's | Visual navigation graph search · Visual Navigation | AUC: 22% higher | Improved |
| Add distributional RL (ablation) | SoRB · Visual Navigation | success rate | Improved |
| Add ensembles of value functions (ablation) | SoRB · Visual Navigation, goals at least 10 steps away | success rate: 10 - 20% | Improved |
| Reduce the replay buffer size by 10x (ablation) | SoRB · Visual Navigation | success rate: 10x | No clear effect |
| Use a [0mMaxDist[0m that is too small or too large (ablation) | SoRB graph construction · Visual Navigation | success rate | Hurt |
| Add off-policy goal relabelling (ablation) | SoRB non-search policy · Goal-reaching | performance | Hurt |
| Add off-policy goal relabelling (ablation) | SoRB search policy · Goal-reaching | performance | No clear effect |
| Use search paths to impose lower bounds on Q-value targets (ablation) | SoRB training · Goal-reaching | learning speed and convergence | Hurt |
| Use TD3-style aggregation of distributional critics (ablation) | SoRB · Goal-reaching | performance | Hurt |
Closest papers by meaning
- Exploration via Planning for Information about the Optimal Trajectory (NeurIPS 2022 · similarity 0.89)
- Memory Based Trajectory-conditioned Policies for Learning from Sparse Rewards (NeurIPS 2020 · similarity 0.89)
- DISCOVER: Automated Curricula for Sparse-Reward Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Diffusion Subgoal Planning for Long-Horizon Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Learning Process Rewards via Visitation Matching for Efficient RL (NeurIPS 2026 · similarity 0.87)
- Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control (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.
