Towards Natural Versatility in Legged Robots through Bio-inspired and Learning-based Frameworks

Quadrupedal legged manipulators (QLMs) exhibit great potential in replacing or aiding humans in hazardous and laborious jobs to mitigate risk to human health. However, these systems are incredibly complex due to their high number of degrees of freedom (DoF) and end-effectors. Hence, for successful deployment in these applications, where the environment is often complex and dynamic, a highly versatile control framework is required. For successful deployment, a framework must generate adaptable and optimal locomotion while undertaking complex manipulation tasks. Animals in the wild demonstrate highly proficient locomotion while interacting with their environment. Yet, existing frameworks fail to achieve both these aspects and struggle to generalise to the chaotic nature of the real-world.

This thesis presents a novel solution to this problem through the development of a bio-inspired hierarchical control framework, working towards achieving the same level of versatility that animals exhibit within QLMs.

This was achieved through instilling the abstracted proficiencies of animal locomotion within various levels of the hierarchical architecture. Animal gait selection strategies were instilled within a DRL gait selection policy, gait memory was embedded within a bio-inspired gait scheduler (BGS), and adaptive motions were realised by a DRL locomotion policy. To facilitate manipulation, whole-body motions, and refine the output of the upstream DRL policies, a whole-body controller (WBC) was developed as the final layer of the hierarchy. Initially, the control modules were tested and validated separately; the locomotion modules achieved zero-shot deployment on challenging terrain, while the WBC completed whole-body manipulation tasks. Once validated on hardware, all control modules were augmented, retrained and unified for loco-manipulation tasks using a QLM, while preserving their respective proficiencies. The resultant framework completed a range of dynamic loco-manipulation tasks in simulation, validating optimal gait selection, whole-body manipulation, and adaptive locomotion simultaneously across a variety of challenging terrains.