Dr. Richard Capraru is currently affiliated with the International Research Center for Neurointelligence (IRCN), the University of Tokyo. His research lies at the intersection of robust autonomous systems, adversarial perception, predictive models, and agency in embodied and AI systems.

He earned his Ph.D. in Electrical and Electronic Engineering from Nanyang Technological University, Singapore in 2026, where he was awarded the Singapore International Graduate Award (SINGA). During his doctoral studies, he conducted research at the Institute for Infocomm Research (I²R), Agency for Science, Technology and Research (A*STAR), Singapore, and was a visiting doctoral researcher at Imperial College London through the NTU–TUM–Imperial Global Fellows Programme.

He earned his B.Eng. in Electrical and Electronic Engineering from University College London in 2021, where he was awarded the Laidlaw Scholarship. He is a Korea University alumnus (member of the Korea University Alumni Association (KUAA)), and was also a visiting student at the Hong Kong University of Science and Technology, Peking University, and the University of Tokyo. His interests include cybersecurity, machine learning, robotics, sensors, and automation. He is an IEEE member.

Email  /  Scholar  /  ResearchGate

Education

Education is a deeply personal journey that shapes not only what we know but also how we think, grow, and learn from others. I have been extremely fortunate and privileged to have studied at the following institutions.

LiDAR Vulnerabilities in Autonomous Vehicle Perception under Rain and Attacks

Minor: Nanotechnology

Academic Visits and Exchanges

Doctoral Research Affiliation

Awards

Research

My research sits at the intersection of robust autonomous systems, adversarial perception, predictive models, sensors, robotics, and embodied intelligence. Broadly, I am interested in how intelligent systems perceive uncertain environments, make structured sense of sensory information, and remain robust under real-world degradation or attack.

Predictive coding and related world-model-based approaches offer a powerful framework for understanding how embodied systems can infer, predict, and adapt. My current work explores these ideas through models of sense of agency, internal representations, and uncertainty in embodied and AI systems.

Autonomous vehicles depend on tightly coupled sensing, learning, and control components. My work in this area examines how vulnerabilities in perception systems, especially LiDAR-based 3D object detection, can be exploited under adverse conditions such as rain, and how these weaknesses affect the safety and robustness of autonomous systems.

My earlier work explored short-range radar sensing for hand gesture recognition, using micro-Doppler signatures to classify subtle motion patterns. This experience contributed to my broader interest in sensing, signal interpretation, and machine perception.

I also worked on Synthetic Aperture Radar (SAR), including image analysis and degradation-related problems in remote sensing. This helped shape my interest in how sensing systems represent and interpret complex real-world signals under imperfect conditions.

Ph.D. Project

As my mentor, Professor Emil Lupu, once reminded me, the purpose of a Ph.D. is not merely to produce a thesis, but to become an independent researcher. I remain deeply grateful for the research training and intellectual guidance that made this work possible.

This thesis studies how adverse weather and adversarial attacks interact to undermine LiDAR-based perception in autonomous vehicles, with a focus on minimal-point ghost-object insertion, object hiding, and defense failure under rain.

Demos / Presentations

Below you can find some demos/presentations of my works.

Publications

Below is a list of my published works, with selected papers highlighted.

This study introduces a rain-aware threat model for LiDAR spoofing, showing how rain can make ghost-object insertion and object-hiding attacks more feasible, stealthy, and effective with a lower attack budget.

This study proposes a geometric, heuristic-based LiDAR attack aimed at real-time ghost-object generation, reducing the data needed for attack construction while preserving physical realizability and improving execution speed.

This study examines how adapting radar and LiDAR detectors to rain can cause catastrophic forgetting, and shows that layer freezing, mixed-weather pre-training, and simulated data augmentation help retain clear-weather performance while improving rain detection.

This paper introduces a genetic algorithm-based LiDAR spoofing attack for rainy conditions that exploits rain-induced degradation and critical detection points to generate effective ghost-object attacks with fewer adversarial points.

This paper studies object-aware point-cloud upsampling for LiDAR detection in rain, using a semi-supervised approach with a small number of labelled simulated objects to improve detection range and accuracy under adverse weather.

This paper uses neural style transfer to generate realistic SAR images corrupted by radio-frequency interference and noise, then applies pre-trained CNNs to classify interference types.

This paper investigates whether very low-cost CW radar can support hand-gesture recognition, developing the required signal-processing setup and showing that CW radar can achieve high accuracy compared with FMCW at much lower cost.

This paper introduces Dop-NET, a shareable micro-Doppler radar database and challenge for dynamic hand-gesture recognition, designed to support benchmarking and training of radar-based classification methods.

Professional Memberships

Miscellanea

Views

@ 2026 Richard Capraru