On Embodied Sensorimotor Intelligence for Mobile Robotic Systems

Embodied intelligence lies at the intersection where artificial intelligence meets physical bodies, and is vividly exemplified by recent advances in mobile robotics enabled by high-performance perception sensors and actuators. This seminar presents recent efforts to develop methodologies and system-level approaches to embodied sensor-actuator intelligence for agile, complex mobile robotic applications, with a particular emphasis on legged robots. 
For dynamic state estimation using multimodal sensors, we develop a new methodology based on continuous-time motion representation that inherently ensures kinematic consistency, representation efficiency, and uncertainty quantification. This framework enables a family of (ego-)motion estimation systems that have been validated in a variety of real-world settings for localization, odometry, and mapping, involving sensors such as inertial measurement unit (IMU), light detection and ranging (LiDAR), ultra-wideband (UWB), and cameras, and have been deployed on wheeled, legged, wearable, and aerial platforms. Furthermore, we introduce sim-to-real reinforcement learning approaches towards reaching motor intelligence in legged locomotion. Building on this foundation, we demonstrate possible ways to incorporate perceptive inputs, enabling robots to navigate through dynamic and unstructured environments. The talk aims to provide balanced insights from both methodological and engineering perspectives, with a strong emphasis on real-world, real-time challenges.

Join at imt.lu/aula1