TechCapsules

AI-summarised brief · reviewed before publication

Microsoft, in collaboration with academic researchers, has developed a new benchmark called GroundedPlanBench to improve robotic decision-making. The benchmark tests AI models' ability to plan tasks while identifying exact locations for each action. Over 1,000 tasks were built from real robot interactions, including direct and open-ended instructions. A training method called Video-to-Spatially Grounded Planning also improved model performance by learning from videos of robots performing tasks, generating over 40,000 grounded plans.