research
Reinforcement learning, embodied AI, and robotic manipulation
My research explores how learning algorithms can enable robots to autonomously acquire complex skills in the real world — reinforcement learning systems that are data-efficient, robust to uncertainty, and deployable on physical robots.
The long-term goal is to build embodied AI systems that learn continuously from interaction, bridging the gap between machine learning theory and real robotic deployment.
research areas
Reinforcement Learning
Algorithms that improve sample efficiency and exploration in environments with sparse rewards and partial observability.
Embodied AI
Learning systems where perception, control, and reasoning are tightly integrated for real-world interaction.
Robotic Manipulation
Learning-based control policies that transfer from simulation to physical robot platforms.
representative work
MOORL — Hybrid Reinforcement Learning
A meta-policy framework that integrates offline datasets with online exploration, improving stability and learning efficiency while addressing distributional shift in offline data.
ReLOAD — Reward-Free Imitation Learning
Intrinsic reward generation using Random Network Distillation, enabling policy learning from unlabeled trajectories without manually designed rewards.
TEACH — Curriculum Learning for RL
A teacher–student framework where temporal variance in value estimates drives goal selection, accelerating learning in sparse-reward tasks.
Active Perception
An RL system that learns to reposition cameras to maximize visual signal quality, improving perception performance under occlusion and noise.
the pipeline
01 / simulation training
02 / perception integration
03 / real-robot deployment
Algorithm design → simulation training → perception integration → real-world robotic deployment.
current direction
Test-time adaptation for diffusion policies
Making robotic manipulation policies robust to real-world uncertainty and distribution shift, combining generative policy models, reinforcement learning, perception-aware control, and sim-to-real transfer.
Research vision
Future autonomous systems must operate in complex physical environments with limited supervision. I want learning algorithms that let robots learn from small amounts of interaction data, adapt to new environments and tasks, integrate perception with control, and stay reliable in the real world.