Research map / Multi-agent systems and games / Human-AI and goal-conditioned
The Surprising Effectiveness of Visual Odometry Techniques for Embodied PointGoal Navigation
Xiaoming Zhao, Harsh Agrawal, Dhruv Batra, Alexander G. Schwing
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 depth to RGB input for VO (RGB-D vs RGB-only) (ablation) | VO model · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 52 -> 61 (Row 1 vs Row 3) | Improved |
| add RGB to depth-only VO input (ablation) | VO model · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 54 -> 61 | Improved |
| add Dropout to last two layers of VO (ablation) | VO RGB-D · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 61 -> 68 (success +7, SPL +5) | Improved |
| replace single forward pass with 10 stochastic Dropout test-time passes (ablation) | VO RGB-D + Dropout · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 68 -> 42 | Hurt |
| use separate per-action models (SepAct) instead of single model (ablation) | VO RGB-D + Dropout · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 68 -> 75 | Improved |
| use 3x wider unified model instead of SepAct (ablation) | VO RGB-D + Dropout · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 75 -> 70 (SPL 56 -> 52) | Hurt |
| expose unified model to action via embedding instead of SepAct (ablation) | VO RGB-D + Dropout · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 75 -> 72 (success and SPL both drop 3 points) | Hurt |
| add geometric invariance losses (ablation) | VO SepAct · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 75 -> 77 (SPL 56 -> 57) | Improved |
| add data augmentation by reversing turn pairs only (ablation) | VO SepAct · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 75 -> 75 | No clear effect |
| add depth discretization with 10 channels (ablation) | VO SepAct+GeoInv · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 77 -> 79 (SPL 57 -> 60) | Improved |
| use 5 instead of 10 depth discretization channels (ablation) | VO full · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 79 -> 74 | Hurt |
| use 20 instead of 10 depth discretization channels (ablation) | VO full · Gibson-4+ val (noisy, no GPS+Compass) | success rate: 79 -> 79 | No clear effect |
Closest papers by meaning
- Is Mapping Necessary for Realistic PointGoal Navigation? (CVPR 2022 · similarity 0.91)
- Learning to Navigate Efficiently and Precisely in Real Environments (CVPR 2024 · similarity 0.88)
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.86)
- Reasoning in Visual Navigation of End-to-end Trained Agents: A Dynamical Systems Approach (CVPR 2025 · similarity 0.85)
- Seeing through Uncertainty: Robust Task-Oriented Optimization in Visual Navigation (NeurIPS 2025 · similarity 0.85)
- Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation (CVPR 2024 · 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.
