Research map / Reinforcement learning / Visual and representation RL
Cross-View Policy Learning for Street Navigation
Ang Li, Huiyi Hu, Piotr Mirowski, Mehrdad Farajtabar
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 cross-view policy learning with aerial-view adaptation (vs single-view agent) | street navigation agent · StreetAir NYU target region | zero-shot reward: 5 -> 29 | Improved |
| add cross-view policy learning with aerial-view adaptation (vs single-view agent) | street navigation agent · StreetAir NYU target region | zero-shot success rate: 10.5% -> 34.5% | Improved |
| add cross-view policy learning (vs single-view) | street navigation agent · StreetAir NYU, reward at 100M steps | reward: 50 -> 190 | Improved |
| add cross-view policy learning (vs single-view) | street navigation agent · StreetAir NYU, reward at 200M steps | reward: 200 -> 280 | Improved |
| add cross-view policy learning (vs single-view) | street navigation agent · StreetAir NYU, convergence | convergence reward: 500 -> 580 | Improved |
| add heading prediction auxiliary task | single-view and cross-view agents · StreetAir NYU, transfer | reward: approximately 450 reward at step 350M | Improved |
| add heading prediction auxiliary task (gap between cross-view and single-view shrinks) | cross-view vs single-view comparison · StreetAir NYU, transfer | reward gap | Mixed |
| add training curriculum (ablation) | single-view agent · StreetAir NYU, transfer | reward at 100M steps: over 50 reward within 100M steps -> 30 less than the one trained with curriculum | Improved |
| add training curriculum (ablation) | cross-view agent · StreetAir NYU, transfer | reward at 200M steps: around 290 at 200M steps | No clear effect |
| add aerial-view adaptation stage (ablation) | cross-view agent · StreetAir NYU, transfer | zero-shot reward and learning speed | Improved |
| add aerial-view adaptation stage (ablation) | cross-view agent · StreetAir NYU, after many training steps | reward | No clear effect |
| add KL policy distillation loss to L2 embedding loss (ablation) | cross-view agent · StreetAir NYU, transfer with curriculum and heading | reward | Improved |
Closest papers by meaning
- Learning to Follow Directions in Street View (AAAI 2020 · similarity 0.91)
- HTNav: A Hybrid Navigation Framework with Tiered Structure for Urban Aerial Vision-and-Language Navigation (CVPR 2026 · similarity 0.89)
- Vision-Language Navigation Policy Learning and Adaptation (TPAMI 2020 · similarity 0.88)
- From Seeing to Experiencing: Scaling Navigation Foundation Models with Reinforcement Learning (ICLR 2026 · similarity 0.87)
- WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation (NeurIPS 2026 · similarity 0.85)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.85)
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.
