Fisher-Preserving Guidance: Training-Free Manifold Constraints for Safe Diffusion Control
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866914614250307584 |
|---|---|
| author | Ren, Hao Bi, Zetong Zeng, Yiming Zheng, Le Li, Zhi Wan, Zhaoliang Qi, Lu Cheng, Hui |
| author_facet | Ren, Hao Bi, Zetong Zeng, Yiming Zheng, Le Li, Zhi Wan, Zhaoliang Qi, Lu Cheng, Hui |
| contents | Diffusion models are effective for waypoint prediction in visual navigation, but standard sampling and test time guidance can produce unreliable or inefficient trajectories when updates drift off the training manifold. We propose Fisher Preserving Guidance with Outer Product Span Projection, a training-free inference method that avoids large Fisher drift associated with off-distribution actions while optimizing a task objective. Our method computes the Fisher-preserving update via a low-rank Jacobian factorization, requiring only a single backward pass per step and enabling real-time use. We further introduce Truncated Fisher Denoising Sensitivity as an uncertainty signal and use it for robust multi-sample action blending. Experiments on toy and realistic navigation benchmarks, including Maze2D with TSDF-based guidance, PushT with official Diffusion Policy weights, and visual navigation in simulation and on real robots, demonstrate consistent improvements in performance over strong diffusion-policy baselines without additional training. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_29937 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Fisher-Preserving Guidance: Training-Free Manifold Constraints for Safe Diffusion Control Ren, Hao Bi, Zetong Zeng, Yiming Zheng, Le Li, Zhi Wan, Zhaoliang Qi, Lu Cheng, Hui Robotics Machine Learning Diffusion models are effective for waypoint prediction in visual navigation, but standard sampling and test time guidance can produce unreliable or inefficient trajectories when updates drift off the training manifold. We propose Fisher Preserving Guidance with Outer Product Span Projection, a training-free inference method that avoids large Fisher drift associated with off-distribution actions while optimizing a task objective. Our method computes the Fisher-preserving update via a low-rank Jacobian factorization, requiring only a single backward pass per step and enabling real-time use. We further introduce Truncated Fisher Denoising Sensitivity as an uncertainty signal and use it for robust multi-sample action blending. Experiments on toy and realistic navigation benchmarks, including Maze2D with TSDF-based guidance, PushT with official Diffusion Policy weights, and visual navigation in simulation and on real robots, demonstrate consistent improvements in performance over strong diffusion-policy baselines without additional training. |
| title | Fisher-Preserving Guidance: Training-Free Manifold Constraints for Safe Diffusion Control |
| topic | Robotics Machine Learning |
| url | https://arxiv.org/abs/2605.29937 |