Seungjae (Jay) Lee

Why Are Web AI Agents More Vulnerable Than Standalone LLMs? A Security Analysis
Jeffrey Yang Fan Chiang*, Seungjae Lee*, Jia-Bin Huang, Furong Huang, Yizheng Chen
(*equal contribution)
+ ICLR 2025 Workshop Building Trust in Language Models and Applications project website / arXiv

Recent studies reveal alarming security flaws in Web AI agents—making them shockingly prone to adversarial attacks, even when built with safety-aligned LLMs. Our research uncovers why these vulnerabilities exist and how they compare to standalone LLMs.

Robot Utility Models: General Policies for Zero-Shot Deployment in New Environments
Haritheja Etukuru, Norihito Naka, Zijin Hu, Seungjae Lee, Julian Mehu, Aaron Edsinger, Chris Paxton, Soumith Chintala, Lerrel Pinto, Nur Muhammad Mahi Shafiullah
ICRA, 2025
+ CoRL 2024 Workshop on Language and Robot Learning, "Oral"
project website/ paper/ github

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.

Behavior Generation with Latent Actions
Seungjae Lee, Yibin Wang, Haritheja Etukuru, H. Jin Kim, Nur Muhammad Mahi Shafiullah, Lerrel Pinto
ICML, 2024 Spotlight (Top: 3.5%)
+ RSS 2024 Workshop SemRob, "Oral spotlights"
+ ICML 2024 Workshop MFM-EAI, "Outstanding Paper Award - Winner"

project website/ arXiv/ github/ 🤗 Lerobot Library

We present Vector-Quantized Behavior Transformer (VQ-BeT), a versatile model for multimodal action prediction, conditional generation, and partial observation handling.

CQM: Curriculum Reinforcement Learning with a Quantized World Model
Seungjae Lee, Daesol Cho, Jonghae Park, H Jin Kim
NeurIPS, 2023 (Acceptance Rate: 26.07%)
arXiv

In this work, we presents a curriculum learning method that uses a quantized world model to automatically generate effective training goals in high-dimensional state spaces.

Diversify & Conquer: Outcome-directed Curriculum RL via Out-of-Distribution Disagreement
Daesol Cho, Seungjae Lee, H Jin Kim
NeurIPS, 2023 (Acceptance Rate: 26.07%)
arXiv

In this work, we develop a method that uses out-of-distribution disagreement to diversify goal selection, enabling curriculum learning from only a few outcome examples.

SNeRL: Semantic-aware Neural Radiance Fields for Reinforcement Learning
Dongseok Shim*, Seungjae Lee*, H Jin Kim
(*equal contribution)
ICML, 2023 (Acceptance Rate: 27.96%)
arXiv / github

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.

Outcome-directed Reinforcement Learning by Uncertainty & Temporal Distance-Aware Curriculum Goal Generation
Daesol Cho*, Seungjae Lee*, H Jin Kim
(*equal contribution)
ICLR, 2023 Spotlight (Top: 5.65%)
arXiv / github

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.

Deep End-to-End Imitation Learning for Missile Guidance with Infrared Images
Seungjae Lee, Jongho Shin, Hyeong-Geun Kim, Daesol Cho, H. Jin Kim
IJCAS, 2023

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: A Graph-Based Approach for Long-Horizon and Sparse Hierarchical Reinforcement Learning
Seungjae Lee, Jigang Kim, Inkyu Jang, H. Jin Kim
NeurIPS, 2022 Oral (Top: 1.76%)
arXiv / github

We present DHRL, a hierarchical reinforcement learning framework that uses a graph-based structure to improve exploration and long-term planning.

Robust and Recursively Feasible Real-Time Trajectory Planning in Unknown Environments
Inkyu Jang, Dongjae Lee, Seungjae Lee, H Jin Kim
IROS, 2021
arXiv

We proposes a trajectory planning algorithm that ensures robust, real-time navigation in unknown environments by maintaining recursive feasibility.