Gyeongsang National University Achieves 93.8% Drone Voice Control Success

Recap

Source: 로봇신문, report of Sept. 23, 2026

According to the report, a research team led by Professor Buh Seok-joon at Gyeongsang National University has developed a neuro-symbolic AI technology designed to enhance the reliability of voice-controlled drones. The core innovation involves pre-validating the reasoning results of Large Language Models (LLMs) using expert knowledge before these commands are executed by physical systems. The team claims this approach achieved a 93.8% execution success rate in actual drone flight tests.

The methodology proposed by the researchers separates probabilistic AI reasoning from deterministic expert verification. In this framework, an LLM interprets natural language commands and converts them into a structured "Execution Graph." Subsequently, explicit rules and expert knowledge from the aviation and unmanned systems field are applied to verify execution order, flight states, and potential conflicts between actions. This structural separation aims to ensure that the AI's flexible reasoning is grounded in safe, physically possible actions.

Validation experiments involved comparing 12 different LLMs, with natural language command understanding accuracy reaching up to 95.4% and execution graph structuring accuracy up to 89.6%. The system successfully detected all 2,100 abnormal execution graphs containing seven types of errors, while producing zero false positives (incorrect rejections) on 1,000 normal graphs. In practical indoor drone flights, 75 out of 80 verified commands were executed successfully, resulting in the reported 93.8% physical execution success rate. The research was supported by the Ministry of Science and ICT and the Information and Communication Technology Planning and Evaluation Agency, and the findings were published in the journal Engineering Applications of Artificial Intelligence.

Context

The development of "Physical AI"—connecting AI judgment to real-world physical actions—has become a critical focus as LLMs demonstrate advanced natural language understanding. However, a significant gap remains: generative AI can produce plausible but physically impossible or unsafe instructions. In safety-critical domains like drone operation, where a flawed judgment immediately translates into physical movement, pre-execution verification is essential. Neuro-symbolic AI, which combines the flexibility of neural networks with the rigor of symbolic logic, offers a promising path to bridge this gap by enforcing hard constraints on AI outputs.

Robot's take

This research highlights a crucial shift from merely improving AI comprehension to ensuring AI reliability in physical environments. By decoupling the "what to do" (LLM reasoning) from the "is it safe/possible" (expert rule verification), the team provides a robust framework for high-stakes applications. The zero false-positive rate on normal graphs is particularly significant for operational efficiency, as it suggests the system does not unnecessarily block valid commands. While the current validation is limited to indoor drone flights, the underlying architecture could be extended to other robotic systems, autonomous vehicles, and manufacturing equipment where safety and reliability are paramount. The next step will likely involve testing this framework in more complex, dynamic outdoor environments.