Research map / Image generation / Fast diffusion sampling
Prior Does Matter: Visual Navigation via Denoising Diffusion Bridge Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NaviBridger, a visual navigation policy built on denoising diffusion bridge models that starts denoising from informative prior actions (rule-based or CVAE-learned) instead of Gaussian noise.
Problem
Diffusion-based visual navigation policies start denoising from Gaussian noise, which differs greatly from the target action distribution, causing redundant denoising steps and harder learning, and sparse effective actions make accurate generation difficult.
Approach
Replace the Gaussian noise starting point of diffusion policies with a prior action sampled from an observation-conditioned source distribution (Gaussian, rule-based parabolic path, or CVAE-learned prior), and denoise to the target action via a denoising diffusion bridge (DDBM), with a theoretical argument that closer source distributions lower the error bound.
Key results
- In Stanford 2D-3D-S indoor basic task, the learning-based prior reaches 92% success with 0.61 collisions versus 86% for NoMaD and 68% for ViNT.
- In CitySim outdoor adaptation task, the learning-based prior achieves 64% success versus 52% for NoMaD and 38% for ViNT.
- Using learning-based prior DDBM, success peaks around 90% at 10 denoising steps, and the DDBM denoising converges in fewer steps than DDPM (NoMaD).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use learning-based prior actions instead of the NoMaD baseline | NaviBridger · Indoor (2D-3D-S), Basic Task | success rate: 86% -> 92% | Helped |
| Use learning-based prior actions instead of the NoMaD baseline | NaviBridger · Indoor (2D-3D-S), Adaptation Task | success rate: 32% -> 88% | Helped |
| Use learning-based prior actions instead of the NoMaD baseline | NaviBridger · Outdoor (Citysim), Basic Task | success rate: 22% -> 44% | Helped |
| Use learning-based prior actions instead of the NoMaD baseline | NaviBridger · Outdoor (Citysim), Adaptation Task | success rate: 52% -> 64% | Helped |
| Use Gaussian prior actions instead of the NoMaD baseline | NaviBridger · Indoor (2D-3D-S), Basic Task | success rate: 86% -> 82% | Hurt |
| Use Gaussian prior actions instead of the NoMaD baseline | NaviBridger · Indoor (2D-3D-S), Adaptation Task | success rate: 32% -> 64% | Helped |
| Use Gaussian prior actions instead of the NoMaD baseline | NaviBridger · Outdoor (Citysim), Basic Task | success rate: 22% -> 34% | Helped |
| Use Gaussian prior actions instead of the NoMaD baseline | NaviBridger · Outdoor (Citysim), Adaptation Task | success rate: 52% -> 50% | Hurt |
| Use rule-based prior actions instead of the NoMaD baseline | NaviBridger · Outdoor (Citysim), Basic Task | success rate: 22% -> 14% | Hurt |
| Use rule-based prior actions instead of the NoMaD baseline | NaviBridger · Indoor (2D-3D-S), Adaptation Task | success rate: 32% -> 44% | Helped |
| Add bridge denoising to Gaussian prior actions (ablation) | NaviBridger · Indoor adaptation task | success rate: 0% -> 64% | Helped |
| Add bridge denoising to rule-based prior actions (ablation) | NaviBridger · Indoor adaptation task | success rate: 8% -> 44% | Helped |
Surprising findings
- The rule-based prior performs worse than baseline in several settings (e.g., outdoor basic task 14% success), showing poorly designed priors can hurt performance.
- DDBM inference is slower per step than DDPM at the same step count, even though fewer steps are needed.
Limitations
- Rule-based prior requires hand-designed rules and lacks generalization across conditions
- Performance degrades in scenarios with high-speed moving objects, requiring faster response or adaptive strategies
- Real-world and simulation evaluation details are limited to the main text; no quantitative real-world results reported in the body
Experimental setup
Datasets: RECON SCAND GoStanford SACSoN
Benchmarks: Stanford 2D-3D-S (indoor simulation) Gazebo CitySim (outdoor simulation) real-world indoor and outdoor robot tests
Metrics: path length (m) collision count success rate MSE to expert actions
Baselines: ViNT NoMaD
Builds on: DDBM NoMaD ViNT EDM
Closest papers by meaning
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.89)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.88)
- DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · similarity 0.88)
- Optimizing Diffusion Noise Can Serve As Universal Motion Priors (CVPR 2024 · similarity 0.88)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.87)
- Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.87)
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.