Seonglae ChoWhy are you applying for this specific programme, and how does it fit with your future plans? (max 250 words)
Interpretable AI, attempting to understand black-box AI's internal procedures, is my genuine interest in Artificial Intelligence (AI). Aware of the potential future threats posed by Humanoid, practical concern motivated me to investigate Explainable AI, specifically Mechanistic Interpretability, to ensure AI functions reliably and operates as intended. If hardware robots experience hallucinations, it could pose a significant threat. This concern motivated me to investigate Explainable AI. My goal is to uncover the underlying mechanisms behind Robotics AI and ensure safe preferences that align with human intentions. As humanoid capabilities exponentially emerge, explainable AI becomes increasingly important, especially in robotics. This passion for interpretable AI guided my research into decision-making within AI. Additionally, my three years of work experience before finishing my Bachelor's degree in the autonomous driving technology sector have reinforced my belief in the necessity of a thorough understanding to ensure the safe and precise operation of AI.
Hallucinations in robotics models pose significant dangers, unlike language models’ hallucinations. Therefore, I aim to prevent such accidents in robotics. King’s College London offers a unique academic environment that aligns with my research interests. The elective modules ‘Agents & Multi-Agent Systems’ and ‘Artificial Intelligence Planning’ align perfectly with my focus, and I am eager to take them. Dr. Yali Du’s recent research on ‘Safe multi-agent reinforcement learning for multi-robot control’ at King’s College London matches my interests exactly. I am excited about devising creative interpretable Reinforcement Learning methods to achieve my academic goals.
How does your experience and education make you a suitable candidate for this programme? (max 250 words)
Spending one year as a software engineer at the company Kakao Mobility, a dominant company in Korea's mobility sector, I recognized my own limitations in contributing to AI. I led the development of the map platform and designed a protocol for an AI feature that bridges machine learning algorithms for integrating multiple sensory data. Furthermore, Kakao Mobility’s testing of a robo-taxi for public use under real-world conditions in Pangyo offered me firsthand exposure to autonomous driving technology. This experience profoundly changed my perspective on Robotics AI, as I observed the challenges of integrating AI models with physical hardware. The stiffness in driving due to the discrete vector of the lane map highlighted the importance of precise data handling in robotics, reinforcing my belief that understanding internal AI systems is essential for investigating and solving operational problems.
Driven by my passion to delve into the mechanisms of AI, I returned to finish my Bachelor’s degree and started focusing on Explainable AI. As an intern at Yonsei University’s Data Intelligence Lab, I conducted research on making AI interpretable. As a result, my work on training the RTSum (Relation-Triple Summarization) model and creating the dataset for the RTSum framework, which split and recombined information at a granular level, led to our paper’s acceptance at the North American Chapter of the Association for Computational Linguistics (NAACL) 2024 conference. This not only enhanced my research skills but also confirmed my belief in the importance of a thorough understanding to effectively direct AI model. Additionally, for my robotics research, I have implemented Reinforcement Learning algorithms like PPO and TD3 from scratch using Gymnasium with MuJoCo benchmark environment.
Please describe an example project in which you have combined writing software and building hardware. (max 250 words)
For two years, I worked as a software engineer at the mobility startup Stryx, gaining an overall understanding of autonomous driving AI data processing. Starting with connecting software and hardware by utilizing firmware, I implemented two projects for the Mobile Mapping System (MMS), namely MMS-Tower and MMS-Vehicle, for multiple vehicle agents to collect data simultaneously. During that project, I combined Geographic Information System (GIS) data with Velodyne LiDAR point cloud data by utilizing internally provided firmware C++ APIs (Application Programming Interfaces). This experience developed a deep understanding of point cloud and image sensor fusion for the mobility data pipeline.
This role enabled me to gain hands-on experience in mounting sensor equipment onto car roof racks and managing hardware setups for a vehicle to effectively collect map data. Later, I took responsibility for the latter part, the vector map generation, providing an overall insight into the data flow. This holistic understanding was instrumental in improving the pipeline, which established me as a key player. My significant contributions led to the delivery of the map for Korea’s first driverless service company, 42Dot, and facilitated Stryx’s successful acquisition by Kakao Mobility.
arxiv.org
https://arxiv.org/pdf/2312.12861