Research map / Reinforcement learning / Offline RL
Improving Diffusion Planners by Self-Supervised Action Gating with Energies
Yuan Lu, Dongqi Han, Yansen Wang, Dongsheng Li
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 SAGE energy gating to the planner | DV* · D4RL MuJoCo average | normalised score: 82.9 -> 84.4 | Improved |
| add SAGE energy gating | DV* · D4RL Kitchen Partial | normalised score: 90.0 -> 96.6 | Improved |
| add SAGE energy gating | DV* · D4RL Kitchen Mixed | normalised score: 73.6 -> 74.5 | Improved |
| add SAGE energy gating | DV* · D4RL AntMaze average | normalised score: 81.6 -> 84.5 | Improved |
| add SAGE energy gating | DV* · D4RL Maze2D average | normalised score: 161.6 -> 163.1 | No clear effect |
| add SAGE gating (inference latency) | DV* with MCSS · wall-clock latency | overhead: 6.8% | Hurt |
| use JEPA action-conditioned latent predictor instead of state-space ridge (ablation) | SAGE energy · Feasibility discrimination D4RL average | AUROC: 0.88 -> 0.98 | Improved |
| use JEPA action-conditioned latent predictor instead of state-space MLP (ablation) | SAGE energy · Feasibility discrimination D4RL average | AUROC: 0.88 -> 0.98 | Improved |
| use pretrained JEPA encoder instead of random latent (ablation) | SAGE energy · Feasibility discrimination D4RL average | AUROC: 0.77 -> 0.98 | Improved |
| lengthen prefix window to K>=20 (ablation) | SAGE · D4RL average | normalised score | Hurt |
| keep all candidates (keep rate 1) instead of 0.8 (ablation) | SAGE · D4RL average | normalised score | Hurt |
| use overly large penalty lambda (ablation) | SAGE · D4RL average | normalised score | Hurt |
Closest papers by meaning
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Efficient Planning with Latent Diffusion (ICLR 2024 · similarity 0.88)
- Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 2025 · similarity 0.88)
- State-Covering Trajectory Stitching for Diffusion Planners (NeurIPS 2025 · similarity 0.88)
- Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · similarity 0.87)
- Latent Plan Transformer for Trajectory Abstraction: Planning as Latent Space Inference (NeurIPS 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.
