BROWSE
Detail
Conference Paper
주행 생성 모델의 학습 효율 가속화를 위한 파라미터 탐색 및 가속화 방법
Cited - time in 
Share 
Share
- Issue Date
- 2026-07
- Citation
- 제어·로봇·시스템학회 학술 대회 (ICROS) 2026, pp.1-2
- Publisher
- 제어·로봇·시스템학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- Recent advancements in Vision-Language Navigation (VLN) have successfully integrated Denoising Diffusion Probabilistic Models (DDPM) to enhance exploration performance. However, the practical deployment of these generative models is severely hindered by prohibitive computational demands and significant I/O bottlenecks in data pipelines, leading to sluggish training convergence. In this work, we propose a comprehensive optimization framework designed to bridge this efficiency gap, achieving a 10-fold acceleration in training speed compared to existing baselines. Our methodology consists of three key technical contributions: KD-Tree based search algorithm for efficient path sampling for DDPM, offline preprocessing to eliminate runtime I/O overhead, and systematic grid search of hyperparameter to maximize CPU-GPU throughput. Empirical results demonstrate that our unified strategy enables rapid model convergence without compromising navigation accuracy, providing a practical guideline for scaling diffusion models in complex robotic tasks.
- Keyword
- Efficient training, VLN, Diffusion policy
- KSP Keywords
- CPU-GPU, Data pipeline, Efficiency gap, Efficient path, Grid Search, KD-Tree, Navigation accuracy, Optimization Framework, Path sampling, Probabilistic models, Search Algorithm(GSA)