Research

What I Do

I aim to build intelligent navigation systems that enable robots to operate autonomously in complex, unstructured environments.

Research Themes

• Foundation Models for Embodied AI
  Developing LLM/VLM-based systems that connect perception, reasoning, and action, enabling robots to interpret complex environments and adapt their decisions to changing conditions.

• World Models for Embodied Intelligence
  Learning action-conditioned models of physical environments and robot behavior to predict future states, anticipate interaction outcomes, and support decision-making under uncertainty.

• Deep Reinforcement Learning for Robot Control
  Developing adaptive control policies for mobile and legged robots, with a focus on sample-efficient learning, hierarchical control, and transfer from simulation to the real world.

• Machine Learning-Augmented Motion Planning
  Combining learned representations, prior experience, and classical planning to generate efficient, dynamically feasible motions in cluttered and challenging environments.

Future Directions

My future research will connect foundation models, predictive world models, and reinforcement learning to enable robots to adapt to unfamiliar environments, reason about the consequences of their actions, and operate reliably under physical constraints.

Hardware Platforms

I work with diverse robot platforms to validate algorithms in both simulation and real-world environments.

Robot	Image	Type	Use Case
Unitree Go1		Quadruped	Visual–LiDAR fusion, RL locomotion
Unitree Go2		Quadruped	VLM navigation, cross-modal perception
Unitree G1		Humanoid	LLM-guided policy learning
Clearpath Jackal		Wheeled UGV	Real-world navigation testing
Clearpath Husky		Wheeled UGV	Outdoor mapping, multi-sensor fusion

Simulation Environments

I design and use multiple simulation platforms for both classical planning and learning-based navigation.

Environment	Image	Description
Gazebo		Classic ROS-based simulator for wheeled robots, supporting costmaps, sensor fusion, and realistic physics.
Isaac Gym		GPU-accelerated simulation for large-scale reinforcement learning and policy optimization.
Isaac Sim		High-fidelity NVIDIA Omniverse simulator for perception, dynamics, and multi-robot coordination.
Newton		Legged locomotion simulation and policy evaluation across diverse terrains.
MuJoCo		Closed-loop simulation of robot dynamics, locomotion control, and autonomous navigation.
mjlab		Unitree Go2 locomotion simulation using mjlab and MuJoCo Warp.
Custom Terrains		Procedurally generated terrains for testing locomotion, stability, and adaptive control.
Self-design Simulation		self-design simulation using C++ for motion planning or multi-goal motion planning.