Research map / Reinforcement learning / Offline RL
BiTrajDiff: Bidirectional Trajectory Generation with Diffusion Models for Offline Reinforcement Learning
Yunpeng Qing, Yixiao Chi, Shuo Chen, Shunyu Liu, Kexuan Zhou, Sixu Lin, Litao Liu, Changqing Zou
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 backward-history generation to forward-only generation (bidirectional) (ablation) | BiTrajDiff + IQL · D4RL locomotion average | normalized return: 80.0 -> 85.5 | Improved |
| add forward-future generation to backward-only generation (bidirectional) (ablation) | BiTrajDiff + IQL · D4RL locomotion average | normalized return: 82.1 -> 85.5 | Improved |
| add backward-history generation to forward-only generation (bidirectional) (ablation) | BiTrajDiff + TD3BC · D4RL locomotion average | normalized return: 78.3 -> 83.5 | Improved |
| add forward-future generation to backward-only generation (bidirectional) (ablation) | BiTrajDiff + TD3BC · D4RL locomotion average | normalized return: 82.7 -> 83.5 | Improved |
| add the greedy reward filter on top of the OOD filter (ablation) | BiTrajDiff + IQL/TD3BC · D4RL learning curves | return | Improved |
| add the OOD (isolation forest) filter (ablation) | BiTrajDiff + IQL/TD3BC · D4RL learning curves | training stability | Improved |
| raise augmented data ratio sigma to 30%-50% (ablation) | BiTrajDiff · walker2d-medium-replay | return | Improved |
| raise augmented data ratio sigma to 100% (ablation) | BiTrajDiff · walker2d-medium-replay | return stability | Hurt |
| increase n of the n-step TD estimator (ablation) | IQL/TD3BC with DA · D4RL n-step TD | return | Hurt |
| replace baseline IQL with BiTrajDiff augmentation | IQL · D4RL Maze average | normalized return: 51.2 -> 69.7 | Improved |
| replace baseline TD3BC with BiTrajDiff augmentation | TD3BC · D4RL Kitchen average | normalized return: 6.9 -> 15.1 | Improved |
| replace baseline DT with BiTrajDiff augmentation | DT · D4RL locomotion average | normalized return: 67.7 -> 71.8 | Improved |
Closest papers by meaning
- DiffStitch: Boosting Offline Reinforcement Learning with Diffusion-based Trajectory Stitching (ICML 2024 · similarity 0.91)
- GTA: Generative Trajectory Augmentation with Guidance for Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
- Temporal Distance-aware Transition Augmentation for Offline Model-based Reinforcement Learning (ICML 2025 · similarity 0.90)
- State-Covering Trajectory Stitching for Diffusion Planners (NeurIPS 2025 · similarity 0.90)
- Structural Information-based Hierarchical Diffusion for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Reasoning with Latent Diffusion in Offline Reinforcement Learning (ICLR 2024 · similarity 0.88)
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.
