Research map / Image generation / Fast diffusion sampling
Masked Generative Policy for Robotic Control
Lipeng Zhuang, Shiyu Fan, Florent P. Audonnet, Yingdong Ru, Edmond S. L. Ho, Gerardo Aragon-Camarasa, Paul Henderson
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 score-based (posterior-confidence) masking with partial refinement instead of re-masking all remaining tokens (ablation) | MGP-Long · Meta-World Hard+Very Hard (10 tasks) | avg success rate: 0.541 -> 0.563 | Improved |
| add online token refinement instead of one-shot full-horizon plan (ablation) | MGP-Long · Meta-World Hard+Very Hard (10 tasks) | avg success rate: 0.340 -> 0.563 | Improved |
| add score-based masking in dynamic environments (ablation) | MGP-Long · Meta-World dynamic (5 tasks) | avg success rate: 0.396 -> 0.436 | Improved |
| use confidence recomputed from the latest observation (ATR) instead of random scores (ablation) | MGP-Long · Meta-World Very Hard (5) | avg success rate: 0.478 -> 0.586 | Improved |
| use ATR recomputed confidence instead of reusing stale confidence (ablation) | MGP-Long · Meta-World Hard (5) | avg success rate: 0.468 -> 0.540 | Improved |
| increase mask-refine iterations from 1 to 2 (ablation) | MGP-Short · Meta-World Hard (5) | avg success rate: 0.396 -> 0.540 | Improved |
| increase refine steps from 2 to 3 (ablation) | MGP-Short · Meta-World Hard (5) | avg success rate | No clear effect |
| increase refine steps from 2 to 3 (ablation) | MGP-Long · Meta-World dynamic (5 tasks) | avg success rate: 0.436 -> 0.418 | Hurt |
| vary codebook size 512, 1024, 2048 (ablation) | MGP-Short · Meta-World Very Hard (5) | avg success rate: 0.534 / 0.538 / 0.522 | No clear effect |
| change tokenization granularity from 4 to 2 or 8 actions per token (ablation) | MGP-Short · Meta-World Very Hard (5) | avg success rate: 0.538 -> 0.526 / 0.514 | Hurt |
| mask bottom 70% of tokens instead of 50% (ablation) | MGP-Long · Meta-World Very Hard (5) | avg success rate: 0.560 -> 0.586 | Improved |
| execution step 12 instead of 4 or 36 (ablation) | MGP-Long · Meta-World Hard (5) | avg success rate: 0.478 / 0.480 -> 0.540 | Improved |
Closest papers by meaning
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.90)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.89)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.89)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · 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.
