SidewalkBench: leaderboard
Visual navigation policies drive a simulated delivery robot along 100 procedurally generated urban sidewalk routes of over 100 m with obstacles and pedestrians; the number is failures (collisions, lane exits and freezes) per 100 m travelled, lower is better. Unlike goal-directed indoor navigation benchmarks (Habitat ObjectNav, VLN-CE), the policy gets no privileged goal and is judged on staying on the sidewalk and handling pedestrians.
Metric: Failures per 100 m (collisions, lane exits and freezes) on 100 long-horizon routes in procedurally generated scenes (lower is better). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | FlowPilot | 1.34 | |
| 2 | MIMIC | 1.41 | |
| 3 | S2E | 2.36 | |
| 4 | MBRA | 7.1 | |
| 5 | NoMaD | 8.7 | |
| 6 | CityWalker | 11.1 | |
| 7 | InternVLA-N1-DualVLN | 12.92 | |
| 8 | ViNT | 25.24 |
Interactive version: theaggregate.ai/benchmark?slug=sidewalkbench · How It Works · Data refreshed daily, snapshot 2026-11-13.