Story perspectives

Microsoft Unveils GroundedPlanBench to Enhance Spatial Reasoning

3/30/2026

Summary

Story summary

Microsoft and academic researchers introduced GroundedPlanBench, a benchmark linking actions to specific locations in images to unify planning and spatial reasoning.
They also introduced Video-to-Spatially Grounded Planning (V2GP), which learns from robot videos to create over 40,000 grounded plans.
Training on this data improved model performance.
The study notes remaining challenges in complex tasks.
It suggests combining this method with predictive models to improve real-time decisions.

1 of 1

Story summary

Microsoft and academic researchers introduced GroundedPlanBench, a benchmark linking actions to specific locations in images to unify planning and spatial reasoning.
They also introduced Video-to-Spatially Grounded Planning (V2GP), which learns from robot videos to create over 40,000 grounded plans.
Training on this data improved model performance.
The study notes remaining challenges in complex tasks.
It suggests combining this method with predictive models to improve real-time decisions.