Research map / Vision-language and multimodal / Robot action and navigation
FM-Steer: Enhance Generalist Policies with Value-Guided Cascaded Denoising
Haoming Song, Delin Qu, Yuanqi Yao, Qizhi Chen, Jiarui Li, Qi Lv, Yiwen Tang, Li Kang, Heng Zhou, Xianqiang Gao, Yuhang Tang, Xiaofan Li and 6 more
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 learned flow verifier with random flow-point selection (ablation) | FM-Steer (π0) · LIBERO | success rate (%): 98.6 -> 84.9 | Hurt |
| add Lite-Flow denoiser to complete remaining flow matching (ablation) | FM-Steer (π0) · LIBERO | success rate (%): 89.8 -> 98.6 | Improved |
| add repeat sampling of multiple flow points (ablation) | FM-Steer (π0) · LIBERO | success rate (%): 93.8 -> 98.6 | Improved |
| add cascaded denoising (noise bound T below 1) so candidates differ in noise level (ablation) | FM-Steer (π0) · LIBERO | success rate (%): 95.9 -> 98.6 | Improved |
| add value-guided sampling and cascaded denoising on top of π0 | π0 · LIBERO overall | success rate (%): 94.2 -> 98.6 | Improved |
| add FM-Steer to π0 | π0 · SimplerEnv WidowX overall | success rate (%): 69.2 -> 79.5 | Improved |
| add FM-Steer to GR00T N1 | GR00T N1 · LIBERO overall | success rate (%): 93.9 -> 97.3 | Improved |
| replace flow verifier with random selection (ablation) | FM-Steer (π0) · SimplerEnv (OXE mixture) | success rate (%): 75.1 -> 61.4 | Hurt |
| remove Lite-Flow denoiser (ablation) | FM-Steer (π0) · real-world robot tasks | success rate: -63% | Hurt |
| replace flow verifier with random selection (ablation) | FM-Steer (π0) · real-world robot tasks | success rate: -78% | Hurt |
| apply test-time verification to autoregressive VLAs (V-GPS, RoboMonkey) vs FM-Steer on flow VLA | OpenVLA · SimplerEnv WidowX overall | success rate (%): 37.7 -> 37.0 (V-GPS), 47.8 (RoboMonkey) | Mixed |
| asynchronous cascaded denoising with Lite-Flow denoiser | FM-Steer · real-world WidowX | control frequency: 4Hz verifier, 90Hz denoiser | Improved |
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- Unified Noise Steering for Efficient Human-Guided VLA Adaptation (NeurIPS 2026 · similarity 0.90)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.89)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.89)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.88)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 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.
