Research

Our research focuses on robust perception and autonomous navigation for safety-critical applications, including firefighting, disaster response, and emergency rescue. We develop learning-based and end-to-end navigation methods that combine multimodal perception with hazard-aware decision-making. By integrating complementary sensing modalities such as thermal imaging and radar, we aim to enable safe, adaptive, and reliable navigation despite poor visibility, extreme heat, dynamic hazards, and degraded sensor performance.

Our research focuses on long-horizon manipulation and agentic robotics, integrating perception, memory, task planning, and policy learning to enable robots to decompose goals, coordinate skills, and execute multi-step tasks. We investigate agents that monitor progress, adapt plans, and recover from failures under partial observability and changing environments. By coupling high-level reasoning with closed-loop control informed by visual and physical feedback, we aim to enable reliable and adaptive manipulation in the real world.

We develop multimodal sensing systems to understand human motion and physical interactions. By combining millimeter-wave radar, vision, wearable sensors, and force measurements, we capture movement, muscle activity, and interaction forces. Through synchronized data acquisition, sensor calibration, and multimodal learning, we build physically grounded representations to support robot learning, human–computer interaction, and rehabilitation.

Our research focuses on generative simulation and robot learning through a real–sim–real pipeline. We reconstruct real-world scenes and use generative models to create diverse virtual environments. Physics-based simulation provides interaction data for policy learning, and we investigate the transfer of learned policies to real robots. We aim to reduce data collection costs and improve generalization to new objects, tasks, environments, and robot platforms.