Saravut Lin
Publications

Research Output

2 peer-reviewed journal articles and 1 conference paper, spanning 3D point cloud segmentation, soft robotic control, and nonlinear observer design.

ABSTRACT

Introduces MiniMarket80, a novel synthetic 3D point cloud dataset of 12,000 cluttered retail scenes designed for single-object segmentation evaluation. Five deep learning architectures (PointNet, PointNet++, DGCNN, PointWeb, Stratified Transformer) are benchmarked, with PointWeb achieving the best real-world performance. A compression study using L1 pruning, INT8 quantisation, and knowledge distillation reduces inference latency to 1.77 s while preserving 0.986 mIoU.

DOI: 10.3390/ai7030096 ↗View Project ↗

ABSTRACT

Presents a cascade PID + Unknown Input Observer (UIO) controller for a soft pneumatic gripper that replaces abrupt solenoid-valve actuation with a smooth, precise pressure profile. Full mathematical modelling couples ideal gas law, DC motor dynamics, and lead screw kinematics. Multi-domain Simscape/Simulink simulation and experimental validation on a 13-component physical plant demonstrate accurate volume tracking between 15 mL and 25 mL.

DOI: 10.3390/robotics12040107 ↗View Project ↗

ABSTRACT

Presents the Unknown Input Nonlinear Observer (UIO) theory applied to estimate the internal volume of a soft pneumatic gripper from measurable pressure and motor signals alone. Validated against experimental data from a BeagleBone Black-based real-time control system. Presented at IEEE HORA 2023.

DOI: 10.1109/HORA58378.2023.10156668 ↗View Project ↗
Certifications
2023

Machine Learning Specialisation

DeepLearning.AI & Stanford University

2023

Machine Learning with Python

IBM

2023

Introduction to Cloud Computing

IBM

2023

The Raspberry Pi Platform and Python Programming for the Raspberry Pi

University of California, Irvine

2022

Modern Robotics, Course 1: Foundations of Robot Motion

Northwestern University

2022

Modern Robotics, Course 2: Robot Kinematics

Northwestern University