Research

Our research focuses on sensing and understanding human motion and the physical interactions between humans and their environments. We develop multimodal sensing systems that integrate millimeter-wave radar, vision, surface electromyography (sEMG), inertial sensing, and force sensing to capture body and hand movements, muscle activity, contact events, and interaction forces. Our work encompasses the entire technical pipeline, from synchronized multimodal data acquisition, sensor calibration, and dataset development to multimodal learning and real-time state estimation. We aim to derive informative representations of human actions, interaction states, and behavioral intent from heterogeneous sensor data.

By moving beyond motion-only representations, we seek to characterize the relationships among movement, muscle activation, physical contact, and force during human motion and dexterous manipulation. These physically grounded representations provide richer and more complete information about human behavior for robot manipulation learning, with broader applications in human–computer interaction and rehabilitation.

Our research focuses on long-horizon and memory-aware mobile manipulation. We develop learning-based systems that integrate perception, memory, navigation, and manipulation to carry out complex, multi-step tasks over extended periods. Our work spans scene understanding, memory representation and retrieval, decision-making, policy learning, and real-world deployment. We aim to enable robots to preserve temporal context, track task progress, adapt to changing operating conditions, and use past observations and interactions to guide future actions.

Long-horizon execution is challenging under incomplete observations and changing scene configurations, where errors can accumulate across subtasks. Reactive policies that rely primarily on current observations often struggle to maintain coherent long-horizon behavior and recover from execution failures. We investigate memory-aware representations and policies that connect past experience and scene understanding with decisions over extended action sequences. By integrating perception, memory, navigation, and manipulation into a closed loop, we seek to develop reliable embodied agents that can operate robustly and adaptively in the real world.

Our research focuses on scalable robot learning through scene and data synthesis. By integrating 3D scene reconstruction, procedural generation, generative models, physics-based simulation, and motion planning, we construct diverse, interactive, and physically consistent virtual environments. Within these environments, we synthesize the visual observations, task specifications, action trajectories, and interaction feedback required for robot learning. Our research spans real-world scene digitization, simulation construction and randomization, task and behavior generation, synthetic-data quality assessment, policy learning, and sim-to-real transfer. We aim to transform limited real-world scenes and robot demonstrations into training environments and interaction data that can be systematically scaled across tasks, objects, and robot embodiments.

This research addresses the high cost of real-world robot data collection, the limited diversity of available environments and tasks, and the scarcity of long-tail interaction data. We investigate the relationships among scene semantics, physical properties, task objectives, and robot behavior. By closing the loop from real-world perception and scene synthesis to large-scale data generation, policy learning, and physical deployment, we seek to improve generalization to unseen objects, environments, and tasks. This research provides scalable data and environment infrastructure for general-purpose robotic manipulation, embodied intelligence, and robot foundation models.