Research map / Reinforcement learning / Visual and representation RL
Lagrangian Perturbation Diffusion Steering: Latent Reinforcement Learning for Generative Policies
Hikmet Simsir, Ozgur S. Oguz
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 |
|---|---|---|---|
| replace unconstrained noise-space steering (DSRL) with Lagrangian trust-region residual perturbation (LP-DS) | frozen diffusion/flow policies · Gym Walker2D | episodic return: roughly 4000 -> approximately 5000 | Improved |
| apply LP-DS to real Franka pick-and-place after simulation adaptation | frozen flow-matching backbone · Real Franka spatial pick-and-place | successful trials: 18/40 -> 33/40 | Improved |
| apply LP-DS to real Franka mug hanging | frozen flow-matching backbone · Real Franka mug hanging | successful trials: 11/20 -> 17/20 | Improved |
| add Lagrangian penalty on perturbation magnitude (ablation) | LP-DS · Adroit Pen | success rate | Improved |
| add explicit hard bound on steering noise (ablation) | LP-DS · Adroit Pen | success rate | No clear effect |
| add both Lagrangian penalty and noise bound (ablation) | LP-DS · Adroit Pen | success rate | Improved |
| add Lagrangian penalty to preserve action entropy (ablation) | LP-DS · Adroit Pen | action entropy | Improved |
| apply residual perturbation in action space (LP-DS-A) instead of noise space (ablation) | LP-DS · Walker2D | return | Hurt |
| increase trust-region bound delta from 0.01/0.05 to 0.1 (ablation) | LP-DS · toy multi-goal navigation | mode coverage | Mixed |
| vary trust-region target delta (0.35, 0.5, 0.66) (ablation) | LP-DS · Hopper, Walker2D, Robomimic Square, Adroit Relocate | reward or success | No clear effect |
| use flow-matching backbone instead of diffusion backbone (ablation) | LP-DS · Hopper-V2 | return | No clear effect |
| use LP-DS instead of DSRL for entropy preservation | noise-space steering · Adroit Pen | k-NN action entropy | Improved |
Closest papers by meaning
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.88)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.88)
- Unified Noise Steering for Efficient Human-Guided VLA Adaptation (NeurIPS 2026 · similarity 0.88)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.88)
- ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning (NeurIPS 2025 · 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.
