Structured memory improves robot planning without changing language models

Researchers at South Korea’s Electronics and Telecommunications Research Institute and the Gwangju Institute of Science and Technology developed SceneQueryPlanner, an open-source system that improves service robot planning through persistent, queryable memory. It outperformed the ReAct baseline across every language model tested, without changing the underlying models.

Rather than feeding a language model an expanding record of observations, the system stores rooms, furniture, objects, attributes and spatial relationships in a 3D scene graph. It retrieves only the information needed for the current planning step and helps robots identify what they must verify before acting. The approach addresses a weakness in conventional systems: relevant observations can become buried in lengthy prompts, encouraging models to substitute plausible assumptions for confirmed environmental facts.

On the team’s AttPlan-Bench household-task benchmark, success with Alibaba’s Qwen 2.5 rose from 44.0% to 68.0%, a gain of 24.0 percentage points. Tasks required reasoning about object properties, appliance operation and activities spanning multiple rooms. Both the code and benchmark are publicly available. Testing was conducted entirely in simulation, however, and no physical deployment results have been published. The next phase will address real-world hardware and noisy, incomplete sensor observations.