Learning Terrain Representation for Legged Robot Localization
In this talk, I will describe a haptic localization system based on a terrain representation learned from the force/torque sensors located at the quadruped robot’s feet. Our approach was twofold, supervised terrain classification and the unsupervised method based on the autoencoder architecture. I will talk about the pros and cons of both approaches. The proposed approach is well-suited for a routine robotized maintenance and inspection application, increasing the robustness of the platform. Finally, I will present an efficient transformer-based representation of the terrain based on force/torque measurements.
Bio
Krzysztof Walas graduated from Poznan University of Technology (PUT) in Poland, receiving MSc in Automatic Control and Robotics. He received (with honours) PhD in Robotics in 2012 for his thesis concerning legged robot locomotion in structured environments. He did his Postdoc at The University of Birmingham, School of Computer Science, Intelligent Robotics Laboratory. He is a recipient of the LIDER project from the National Center for Research and Development. He was a PI in the THING project on subterranean legged locomotion within Horizon 2020. He is PI in REMODEL project on deformable object manipulation within Horizon 2020. Currently, he is an Assistant Professor at the Institute of Robotics and Machine Intelligence at PUT, Poland. His research interests are related to robotic perception for physical interaction applied both to walking and grasping tasks.