AI Tech Flash ── 2026-10-27
ETRI and GIST build robot AI that queries its memory
Robot AI searches its own 3D memory; complex-task success rises
What changes
A robot can look up the location, material and condition of objects it has seen before, and use that to decide its actions. According to ETRI, it checks stored information instead of guessing an object's properties.
For example, when asked to put glass cups and plastic cups in different places, it confirms each cup's material from memory before moving it.
Why it matters
Robot AI built on large language models (LLMs, AI trained on large amounts of text) has had weaknesses. According to Chosunbiz, the more environmental information there is, the harder it is to find what is needed. When information is lacking, the AI can infer incorrectly.
Accuracy drops in particular on complex tasks that require telling apart an object's material, operating condition or spatial relationships. The new technology targets this problem.
What it is
SceneQueryPlanner is an AI agent technology that combines a robot's observation of its environment with queries to its memory. ETRI, the Electronics and Telecommunications Research Institute of South Korea, announced it.
Rooms, furniture, and objects' positions, materials and states are stored as a 3D scene graph. A scene graph is a structured record of the objects in a space and how they relate. The robot updates it each time it observes something new.
Who and where
ETRI developed it jointly with GIST, the Gwangju Institute of Science and Technology, according to reports of the institutes' announcement.
The researchers expect it to be applied to physical AI, meaning robots that perceive, reason and act autonomously in complex real-world settings. Expected uses include home-care and service robots.
Chosunbiz reports that the code for the technology and the evaluation framework has been released. The findings were presented at EMNLP 2026 Findings.
Why it was built
Robots need to know not only where objects are but also what they are made of and what condition they are in. Herald Business calls this a far more demanding challenge.
According to Chosunbiz, earlier LLM-based robot AI struggles to find what it needs when information is abundant, and makes wrong inferences when information is missing. The technology was built to address both problems.
How it works
SceneQueryPlanner combines two processes. In environment grounding, the robot moves through its surroundings to gather information. In memory grounding, it selectively queries its stored memory for only the relevant details.
According to Herald Business, this lets the robot judge both what task to perform and what extra information to seek before acting. In the cup-sorting example, it examines the cups on a table, checks their material in memory, and moves each to its designated place.
Key points
- ETRI and GIST developed SceneQueryPlanner, a robot AI that queries its own stored 3D memory.
- In ETRI's announcement, success rose from 46.8% to 69.6% with Llama 3.1 and from 44.0% to 68.0% with Qwen 2.5.
- Chosunbiz reports that the results come from an evaluation environment the research team built, and real-robot deployment is future work.
Conclusion
ETRI and GIST announced a technology in which a robot decides its actions by querying a stored 3D memory on its own. In ETRI's announcement, complex-task success rose by more than 20 percentage points with Llama 3.1 and Qwen 2.5. According to Chosunbiz, the results come from an evaluation environment built by the research team, and handling incomplete camera and sensor data and moving to real robots remain future work.
Quotes and sources
“The key innovation is that the robot can retrieve what it needs from stored memory at the right moment, without being prompted.”── 헤럴드경제, 2026-10-26
“In tests using Meta's AI model Llama 3.1, the task success rate rose from 46.8 percent to 69.6 percent, a gain of 22.8 percentage points.”── 헤럴드경제, 2026-10-26
“However, these results were verified in an evaluation environment built by the research team.”── Chosunbiz, 2026-10-26
“The team released the code for the related technology and evaluation framework.”── Chosunbiz, 2026-10-26
Source articles
Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.