Research map / Reinforcement learning / Imitation learning
Much Ado About Noising: Dispelling the Myths of Generative Robotic Control
Chaoyi Pan, Giri Anantharaman, Nai-Chieh Huang, Claire Jin, Daniel Pfrommer, Chenyang Yuan, Frank Permenter, Guannan Qu, Nicholas Matthew Boffi, Guanya Shi, Max Simchowitz
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 stochasticity injection and supervised iterative computation (MIP) to regression (ablation) | regression policy · Tool-Hang (state) | success rate: 0.54 -> 0.80 | Improved |
| add iterative computation alone (residual regression) (ablation) | regression policy · Tool-Hang (state) | success rate: Regression << RR 0.54 | Improved |
| increase integration steps of straight flow from 1 to 9 (ablation) | straight flow · Tool-Hang (state) | success rate: 0.54 -> 0.52 | No clear effect |
| supervise intermediate steps (ablation) | MIP · Tool-Hang (state) | success rate: 0.42 -> 0.80 | Improved |
| MIP vs flow with fewer integration steps (2 vs 9 NFEs) | MIP · LIBERO 10 | success rate: 81.6 -> 82.2 | No clear effect |
| use l1 instead of l2 loss for regression (ablation) | regression policy (Sudeep-DiT) · Tool-Hang ph (state) | success rate: 0.50/0.37 -> 0.76/0.65 | Improved |
| use l1 instead of l2 loss for l1 on LIBERO regression (ablation) | pi0 regression · LIBERO 10 | success rate: 78.0 -> 62.4 | Hurt |
| perturb observations during training (ablation) | MIP · Transport / Tool-Hang (state) | success rate: MIP 0.92/0.85 -> MIP-Obs 0.13/0.08 (Tool-Hang, Chi-Transformer) | Hurt |
| use flow instead of regression | policy · Tool-Hang (state) | policy Lipschitz constant: 1.71 -> 1.41 | Improved |
| MIP vs consistency trajectory model at 2 steps | MIP · Tool-Hang (state) | success rate: 0.56/0.26 -> 0.92/0.88 | Improved |
| train on lower-quality (delayed and noisy) data, flow vs regression | Chi-UNet · Tool-Hang data quality | success rate: regression 0.70/0.63, flow 0.76/0.68 | No clear effect |
Closest papers by meaning
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.91)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.91)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.90)
- Make-An-Agent: A Generalizable Policy Network Generator with Behavior-Prompted Diffusion (NeurIPS 2024 · similarity 0.89)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 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.
