GroundedPlanBench: Spatially grounded long-horizon task planning for robot manipulation
Vision-language models still can't tell robots where to move. Microsoft Research just proved it.
Why it matters
Academic research identifying a fundamental gap in VLM-based robot planning—the inability to ground spatial reasoning in real environments. This matters for founders building embodied AI and investors evaluating whether current VLMs are ready for physical-world deployment.
The key facts
9 to knowVision-language models struggle with spatial grounding in robot manipulation tasks
Current two-step approach (VLM planning + separate action translation) frequently fails
GroundedPlanBench benchmark from Microsoft Research
Research focuses on long-horizon task planning for robot manipulation
Published by Microsoft Research (established credibility)
Problem: VLMs struggle with spatial reasoning + action sequencing in robot manipulation
Current approach limitation: Two-step pipeline (natural language plan → executable actions) breaks down in practice
Focus: Long-horizon task planning with spatial grounding requirements
Source: Microsoft Research official publication
Go to the source
Microsoft Researchmicrosoft.com
Publisher excerpt: Vision-language models (VLMs) use images and text to plan robot actions, but they still struggle to decide what actions to take and where to take them. Most systems split these decisions into two steps: a VLM generates a plan in natural language, and a separate model translates it into executable…