BROWSE
Detail
Conference Paper
Attention 분석을 활용한 Pick-and-Place 과제 실패 사례 분석
Cited - time in 
Share 
Share
- Issue Date
- 2026-07
- Citation
- 제어·로봇·시스템학회 학술 대회 (ICROS) 2026, pp.227-228
- Publisher
- 제어·로봇·시스템학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- This study analyzes failure cases of a SmolVLA-based pick-and-place task using the SO-101 robot through attention analysis. We compare Model A, trained with 750 teleoperation episodes, and Model B, trained with 1,050 episodes after adding post-approach pick-to-place data. Although Model B used more data, its real-world success rate did not improve. We therefore define image-mass and object-ratio from action-to-image cross-attention with SAM2-based target masks to examine this failure. The results show that Model B shifted the camera-wise attention distribution but did not improve target object grounding.
- Keyword
- SmolVLA, Pick-and-Place, Cross-Attention, Object Grounding, Failure Analysis
- KSP Keywords
- Attention distribution, Failure analysis, Model B, Object grounding, Pick-and-place, Real-world, Success rate, target object