SceneQueryPlanner: 3D-graph robot memory lifts task success by 24 points

South Korea's ETRI and GIST unveiled SceneQueryPlanner, an open-source system that stores robot observations in a 3D scene graph and queries only task-relevant attributes instead of feeding a growing text context to the LLM. On the new AttPlan-Bench, success rose from 46.8% to 69.6% with LLaMA 3.1 and from 44.0% to 68.0% with Qwen 2.5.

• LLaMA 3.1 success rose 22.8 points, Qwen 2.5 gained 24.0 points
• On GPT-4o, attribute-based composite tasks improved by 17.2 points
• Code and AttPlan-Bench released Oct. 26; results shown at EMNLP 2026
• The underlying language model was unchanged — the memory architecture drives gains