Research map / Reinforcement learning / Visual and representation RL
D²PPO: Diffusion Policy Policy Optimization with Dispersive Loss
Guowei Zou, Weibing Li, Hejun Wu, Yukun Qian, Yuhang Wang, Haitao Wang
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 dispersive loss regularization during diffusion policy pre-training | DPPO · RoboMimic Lift to Transport (pre-training) | success rate: +3.4% (Lift) to +36.2% (Transport); average 22.7% | Improved |
| add dispersive pre-training before PPO fine-tuning | DPPO · RoboMimic Transport (fine-tuned) | success rate: 0.60 -> 0.87 | Improved |
| add dispersive pre-training before PPO fine-tuning | DPPO · RoboMimic Square (fine-tuned) | success rate: 0.83 -> 0.89 | Improved |
| add dispersive pre-training before PPO fine-tuning | DPPO · RoboMimic Lift and Can (fine-tuned) | success rate: 1.00 -> 1.00 | No clear effect |
| add dispersive pre-training before PPO fine-tuning | DPPO · Real robot Transport | success rate: 45% -> 70% | Improved |
| set dispersive loss coefficient lambda to 0.5 (ablation) | D2PPO pre-training · RoboMimic Square (pre-training) | success rate: 14.3% improvement over baseline | Improved |
| set dispersive loss coefficient lambda to 0.3 (ablation) | D2PPO pre-training · RoboMimic Square (pre-training) | success rate | Hurt |
| set dispersive loss coefficient above 0.5 (ablation) | D2PPO pre-training · RoboMimic Square (pre-training) | success rate | Hurt |
| apply dispersive loss in early layers versus late layers (ablation) | D2PPO · RoboMimic simple vs complex tasks | success rate | Mixed |
| use L2 dispersive loss on middle layers versus other variants (ablation) | D2PPO · RoboMimic pre-training | stability and variance | Improved |
Closest papers by meaning
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.92)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.91)
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.90)
- AffordDP: Generalizable Diffusion Policy with Transferable Affordance (CVPR 2025 · similarity 0.89)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.89)
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · 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.
