NUST Institutions Library Catalogue

Quadrupedal robots have gained significant research interest due to their ability to
achieve agile and stable locomotion over complex terrains. Such locomotion can be
achieved by combining various gaits, however, simply changing robot gaits does not
guarantee robust and stable behavior. To ensure stable robot locomotion, gaits must
be seamlessly blended. Current methods of gait transition include model-based, mainly
Model Predictive Control (MPC), approaches, which are limited by the use of handengineered gaits; Reinforcement Learning (RL)-based methods, which address these
limitations but require extensive training; and hybrid methods that combine multiple
controllers but still experience abrupt gait timing changes. This thesis introduces a
novel RL-MPC hybrid control framework that addresses the controllers’ shortcomings
in the current literature. The proposed controller incorporates a feature extractor module that extracts features from the robot terrain and state. The novel framework also
introduces a gait timing correction step to smooth out gait transitions. The proposed
framework was tested on a randomly generated rough terrain, where the robot efficiently traversed and transitioned between gaits while maintaining accurate command
velocity. Testing the effectiveness of the contact timing correction step revealed that the
locomotion produced by the controller without contact timing correction was jerky and
unstable on rough terrain. The proposed framework also outperforms a state-of-the-art
method in gait transitioning, resulting in smoother and more stable locomotion.