Abstract:
With recent performance improvements, legged robots will soon enter our lives to stay. Various control algorithms are already employed in deploying existing robots, and many more algorithms are still in the making. Safety concerns during the operation of legged robots must be addressed to enable their widespread use and ease their development. Especially machine learning- based control methods would benefit from model-based constraints to improve their safety. This thesis presents a modular safety filter to improve safety, i.e., reduce the chance of a fall of a legged robot. The availability of a robot capable of locomotion is assumed, i.e., a nominal controller exists. During locomotion, terrain properties around the robot are estimated through machine learning which uses a minimal set of proprioceptive signals. A novel deep-learning model utilizing an efficient transformer architecture is used for terrain estimation. A quadratic program combines the terrain estimations with inverse dynamics and a novel control barrier function constraint to filter and certify nominal control signals. The result is an optimal controller that acts as a filter and the filtered control signal allows the safe locomotion of the robot. The resulting approach is general and could be transferred with low effort to any other legged system.
