Benchmark Radar360CityArena
360CityArena evaluates embodied agents in a photorealistic virtual urban environment built from 360-degree video of Tokyo's Akihabara district. It includes 175 tasks across environment understanding, path reasoning, and spatial reasoning, testing localization, landmark search, path planning, and relational spatial reasoning. Scoring uses mean evaluation score with varied success criteria.
- Released
- 2026-08-09
- Readiness
- Runnable
- Primary field
- Robotics & Autonomous Systems
Why it matters
Existing outdoor navigation benchmarks lack photorealism or complexity, leaving a gap between simulated and real-world urban conditions. This benchmark provides a repeatable protocol with released code, tasks, and evaluation scripts, enabling comparative assessment of embodied agents' city-scale navigation and reasoning abilities.
Motivation
We present 360CityArena, a benchmark for evaluating the urban exploration capabilities of embodied agents within a photorealistic environment constructed from 360-degree videos.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.