PhD student working on robot learning, world models, embodied AI, and visual model predictive control.
My research focuses on learning systems that enable embodied agents to plan and act under partial observability, visual occlusion, and model uncertainty.
My interest in intelligent systems started before university through Go. In high school, I played Go competitively and became a provincial champion. Like many Go players, I was amazed by AlphaGo’s superhuman performance. Yet one detail left a stronger impression on me than the victory itself: even when AlphaGo could decide where to play, a human still had to physically place the stones on the board.
That contrast shaped how I thought about intelligence. Strategic reasoning could already exceed human experts in a clean symbolic world, while acting in the physical world remained much harder. A real robot must perceive incomplete surroundings, model uncertain dynamics, move through hardware constraints, and recover from noisy feedback.
This belief guided my academic path: mechanical engineering in my bachelor’s degree to understand physical systems, robotics in my master’s degree to connect perception, planning, and control, and electrical engineering in my PhD to study learning-based visual control and embodied AI.
Learning systems that perceive, plan, and act in the physical world under uncertainty and deployment constraints.
Latent dynamics, world action models, long-horizon imagination, model-based RL, and visual MPC.
Partial observability, visual occlusion, uncertainty-aware returns, ensemble calibration, and sim-to-real transfer.
The three main project pages are now organized as standalone visual paper pages: ELVIS, AREPO, and LFD-DRF.
RoBetArme is the physical robotics context that shaped my current research direction. The project targets robotic material deposition in challenging construction environments, where perception, control, material flow, and hardware constraints must be handled together.
This project motivated my focus on robot learning methods that connect visual perception, uncertainty, simulation, planning, and real-world validation.
ELVIS is a latent model predictive control framework for long-horizon visual control. It plans inside a Dreamer-style recurrent state space world model and uses Gaussian-mixture MPPI to maintain multiple coherent action hypotheses over long horizons.
An ensemble-calibrated uncertainty-aware return stabilizes deep imagination and aligns actor-critic learning with trajectory optimization.
AREPO studies uncertainty-aware robot learning for visual control tasks with severe partial observability. The method combines visual representation learning, uncertainty-guided reinforcement learning, and ensemble-based policy learning to improve robustness under occluded observations and sim-to-real shift.
This work was accepted by IEEE Robotics and Automation Letters in 2025 and presented at CASE 2025.
This work studies imitation learning for autonomous driving in safety-critical scenarios. It uses Driver’s Risk Field agents to generate realistic interactive traffic scenarios and improve the safety and efficiency of learned motion planning policies.