VLM agent hits 71% map score in 12 minutes on SpatialScout benchmark

NASA JPL reported the first real-world run of a vision-language model agent on the SpatialScout benchmark: a TurtleBot 4 robot mapped 71% of the environment in 12 minutes with 0.18 m positional error and 22 W power draw, 18 points above the static baseline.

• Map coverage reached 71% versus a 50% target in 12 minutes, not 15
• Positional RMSE was 0.18 m against a 0.30 m target
• Power draw fell to 22 W from 45 W in the baseline setup
• 210 papers cited SpatialScout by November 2026