SpatialAfford framework enhances compact VLMs for precise affordance grounding

Researchers have developed SpatialAfford, a novel two-stage framework designed to improve the affordance grounding capabilities of compact vision-language models (VLMs). This framework first aligns the model's attention to the correct affordance region using Spatial Attention Alignment (SAA) and then refines coordinate prediction with Spatial-Aware GRPO. By explicitly guiding the model's visual focus before it generates coordinates, SpatialAfford enhances spatial reasoning for tasks like grasping or pressing specific object parts. Experiments on benchmarks such as ShareRobot-Bench, ReasonAff, and PartAfford show that SpatialAfford consistently improves performance, with a smaller 4B model surpassing larger 7B+ baselines. AI

IMPACT Improves the precision of compact VLMs in embodied AI tasks by enhancing their ability to identify specific interaction regions.

RANK_REASON Research paper detailing a new framework for improving vision-language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Hugging Face
• PartAfford
• ReasonAff
• ShareRobot-Bench
• SpatialAfford
• Spatial Attention Alignment
• Spatial-Aware GRPO

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →