Self-Supervised Learning for Online Adaptation of Motion Models for Foresighted Autonomous Navigation of Quadruped Robots

• Applicant: Professor Dr. Jörg-Dieter Stückler
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 556093512

Autonomous navigation on diverse and uneven terrain such as asphalt, mounds, or sand, is an important problem in robotics with a plethora of applications in, for example, autonomous driving, outdoor service robotics, or construction and agricultural robotics. Enabling mobile robots with adaptive and efficient navigation behavior on such terrains is an active research field. This requires planning of robust, time- and control-cost-efficient paths in the environment. It is infeasible, however, to describe a large set of possible variations of terrain and robot properties (e.g., payload, battery level) manually by engineers. Instead, methods are desirable which facilitate adaptability of the robot system to the multitude of properties by learning. Mobile robots need to be capable of learning the effects of control inputs in their environment, to predict the properties of terrain such as inclination, slip, or damping using sensors and experience, and to use this information for navigation planning. In this project, we will research novel methods for self-supervised learning of motion models for quadruped robots in environments with diverse and uneven terrain. We assume that basic locomotion control for the given terrains is available, and that the resulting motion of the robot exhibits different properties in terms of effective acceleration and velocity depending on the terrain and robot properties. We will develop methods for learning motion models which capture the behavior of the robot depending on control inputs on diverse terrain. To this end, we will develop novel probabilistic context-conditional motion models which allow for inferring terrain and robot properties as context variables from interactions and sensor data (e.g., visual, auditive, inertial). The learned models will in turn be used for foresighted path planning of confidently assessable and control-cost-efficient paths by model-predictive control. The project will provide novel findings in learning methods for autonomous mobile robot navigation. This can enable novel applications for intelligent robot systems that require foresighted adaptive, robust, and energy-efficient navigation on diverse and uneven terrain. Moreover, our insights can inspire novel research in the broader context of self-supervised learning for online adaptation in robotics or in related disciplines such as cognitive science.

• DFG Programme: Research Grants