KJCKorea
Journal Central

Path planning methods play a major role (for example, collision avoidance and time/cost reduction) in unmanned aerial vehicles (UAV) based services. Path planning methods for ground robots cannot be directly applied to UAVs since they only consider 2D free space. A UAV should be able to fly over or fly around obstacles to reduce flight distance. The proposed path planning method is able to compute a ‘fly-over’ path with consideration of location, width and height of obstacles. The proposed path planning method is more efficient than existing Wavefront method when the obstacle height is less than half the width of the obstacle.

This paper presents a personalized programming education system that integrates a GPT-4-based role-play simulation and an adaptive rubric-driven automated evaluation engine. The system employs prompt engineering to guide GPT-4 in dual roles—as a conversational tutor and as an evaluator— enabling dynamic learner modeling and self-explanation through contextual dialogue. A total of 450 question–answer interactions were collected across three instructional scenarios and three learner personas. Experimental results demonstrated improvements in learning outcomes through repeated sessions, with rubric score consistency within ±0.5 points and a 91.3% agreement rate between AI-generated and human evaluator scores. These findings confirm the system's effectiveness in delivering real-time personalized feedback and formative assessment, and suggest practical applications for AI-driven education.

Large Language Models have advanced natural language understanding, yet remain limited in handling complex multi-hop queries requiring integration across multiple documents. Traditional Retrieval-Augmented Generation adopts a linear query-retrieval-generation pipeline, which often causes error propagation, incomplete evidence coverage, and reduced reliability. To overcome these issues, this study proposes a Tree-of-Retrieval based RAG (ToR-RAG). ToR-RAG decomposes queries into binary sub-queries via LLM prompting, performs retrieval and partial answer generation at each branch, and evaluates outputs using an LLM-as-a-Judge module. Branches below a quality threshold are pruned, ensuring efficiency. An MMR-based retrieval strategy (λ=0.75) with top-k=5 selection balances relevance and diversity. Experiments on the MultiHop-RAG dataset show that ToR-RAG improves Exact Match by +6.42 and F1 by +6.04 compared to Non-RAG, Native-RAG, and CoR-RAG. Performance peaked at Depth=3, while Depth=4 caused degradation from excessive branching and token usage. These results demonstrate that ToR-RAG enhances both accuracy and reliability in multi-hop reasoning, suggesting applicability in domains such as policy analysis, medical decision-making, and financial risk assessment.