Seungjae (Jay) Lee

I am first-year Ph.D. student at UMD Department of Computer Science, co-advised by professors Furong Huang and Jia-Bin Huang.

Prior to UMD, I had my Masters Degree in Department of Aerospace Engineering at SNU. advised by Prof. H. Jin Kim. I also spent time at Generalizable Robotics and AI Lab (GRAIL) at NYU, working closely with Professor Prof. Lerrel Pinto. I worked on enhancing the data efficiency of Reinforcement Learning (RL) and Imitation Learning (IL) systems and applied them to various decision-making scenarios, including real-world robots.

Before that, I received Bachelor's degrees in Mechanical and Aerospace Engineering at SNU.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

News

• 05/2024: A paper on multi-modal behavior generation for robot agents were accepted to ICML 2024 (Spotlight).
• 02/2024: Graduated from the master's program at Seoul National University (Aerospace Engineering).
• 12/2023: Presented two papers on curriculum learning for robot agents at NeurIPS 2023.
• 07/2023: Presented a paper on 3D representation learning for robot agents at ICML 2023.
• 05/2023: Presented a paper on exploration for RL at ICLR 2023 (Spotlight).

Research

My research interest is understanding the interaction between agents and environments, and devising data-efficient decision-making (or robot learning) algorithms, especially in the field of reinforcement learning (RL).

Robot Utility Models (RUMs) is a simple method to build zero-shot robot policies that can solve useful tasks in completely new homes without any additional training often at 90%+ success rate.

In this work, we present Vector-Quantized Behavior Transformer (VQ-BeT), a versatile model for behavior generation that handles multimodal action prediction, conditional generation, and partial observations. Across seven environments including simulated manipulation, autonomous driving, and robotics, VQ-BeT improves on state-of-the-art models such as BeT and Diffusion Policies.

Previous approaches often face challenges when they generate curriculum goals in a high-dimensional space. To alleviate it, we propose a novel curriculum method for agents that automatically defines the semantic goal space, and suggests curriculum goals over it.

Unlike previous curriculum learning methods, D2C requires only a few examples of desired outcomes and works in any environment, regardless of its geometry or the distribution of the desired outcome examples.

We present Semantic-aware Neural Radiance Fields for Reinforcement Learning (SNeRL), which jointly optimizes semantic-aware neural radiance fields (NeRF) with a convolutional encoder to learn 3D-aware neural implicit representation from multi-view images.

We propose an uncertainty & temporal distance-aware curriculum goal generation method for the outcome-directed RL via solving a bipartite matching problem. It can provide precisely calibrated guidance of the curriculum to the desired outcome states.

We propose an end-to-end missile guidance algorithm from raw infrared image pixels by imitating a conventional guidance law which leverages privileged data.

DHRL provides a freely stretchable high-level action interval, which facilitates longer temporal abstraction and faster training in complex tasks.

We propose a real-time robust planner that recursively guarantees persistent feasibility without any need of braking.

Projects

with HD Hyundai Heavy Industries Co., Ltd.
Apr. 2023 - Mar. 2024

Awards and Achievements

Academic Services

Professional Experience

Source code credit to Dr. Jon Barron