Benchmark RadarGST-Bench
GST-Bench is a VQA benchmark for global spatial intelligence in video understanding, covering synthetic videos and human-verified questions. It evaluates VLMs' ability to infer spatial relations from novel viewpoints and map egocentric observations to top-down views.
- Released
- 2026-08-06
- Readiness
- Paper only
- Primary field
- Robotics & Autonomous Systems
Why it matters
Existing video benchmarks focus on local spatial perception, while GST-Bench targets global spatial awareness over long-horizon videos, addressing a gap in evaluating embodied agents' spatial intelligence. It provides a scoring contract to compare VLMs and highlights a significant performance gap between models and humans.
Motivation
Spatial intelligence is fundamental to embodied agents, yet existing benchmarks focus on local spatial perception from single or few viewpoints, overlooking global spatial awareness over continuous, long-horizon visual streams.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.