Research map / Reinforcement learning / Visual and representation RL
Hindsight Experience Replay
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 HER replay to DDPG | DDPG · Pushing, sliding, and pick-and-place | task success | Helped |
| Add HER replay to DQN | DQN · Bit-flipping | largest n solved: n ≤13 -> n up to 50 | Helped |
| Add HER replay (ablation) | DDPG with one fixed goal · Pushing, sliding, and pick-and-place with a single goal | task success | Helped |
| Train with multiple goals rather than one fixed goal (ablation) | DDPG+HER · Pushing, sliding, and pick-and-place | learning speed | Helped |
| Replace sparse binary rewards with the tested shaped rewards (ablation) | DDPG · Pushing, sliding, and pick-and-place | task success | Hurt |
| Replace sparse binary rewards with the tested shaped rewards (ablation) | DDPG+HER · Pushing, sliding, and pick-and-place | task success | Hurt |
| Replay goals sampled from the same episode after the transition (future strategy) instead of the final-state goal (ablation) | DDPG+HER · Sliding | test performance | Helped |
| Use random goals encountered across the whole training procedure for replay (ablation) | DDPG+HER · Pushing and pick-and-place | test performance | Hurt |
| Increase the number of replay goals above 8 (ablation) | DDPG+HER with the future strategy · Pushing, sliding, and pick-and-place | test performance: 8 | Hurt |
| Add gaussian observation noise (std=1cm) during retraining (ablation) | DDPG+HER pick-and-place policy · Pick-and-place, physical robot | success rate: 2 out of 5 -> 5/5 | Helped |
| Use HER rather than count-based exploration | DDPG · Pushing, sliding, and pick-and-place | task success | Helped |
Closest papers by meaning
- DHER: Hindsight Experience Replay for Dynamic Goals (ICLR 2019 · similarity 0.92)
- Cluster-Based Sampling in Hindsight Experience Replay for Robotic Tasks (Student Abstract) (AAAI 2024 · similarity 0.89)
- Goal-Conditioned On-Policy Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- PIPER: Primitive-Informed Preference-based Hierarchical Reinforcement Learning via Hindsight Relabeling (ICML 2024 · similarity 0.87)
- Learning Process Rewards via Visitation Matching for Efficient RL (NeurIPS 2026 · similarity 0.87)
- Provable Interactive Learning with Hindsight Instruction Feedback (ICML 2024 · similarity 0.86)
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.