Quadruped robots have seen rapid growth over the past decade, after a major breakthrough established with MIT Cheetah [Seok13, Katz19]. They provide robust mobility and remarkable agility in unstructured environments. They are almost systematically based on a similar rigid architecture, equipped with quasi direct drive pancake motors all positioned in the torso to minimize leg inertia [Wensing17]. Such rigid structures favor model accuracy and control tractability [Wensing23], but they also limit impact absorption and prevent the robot from exploiting local flexibility to improve mobility. Although quadrupeds are mature, their performance is plateauing far below that of biological counterparts.
Many alternative mechatronic solutions, such as elastic actuators (SEA [Pratt95, Verstraten16], PEA [Mettin10, Ding24], VSA [Vanderborght13]), tensegrity mechanisms [Motro03, Vimbert26], flexible links [Xu22], soft and passive joints [Collins05, Badri-Spröwitz13], have already been proposed in the literature to enhance energy efficiency, improve shock absorption, provide intrinsic compliance, or reduce the robot mass, assembly constraints, and fabrication cost. Embedding such propositions into an active trunk/spine [Ijspeert07, Khoramshahi13] or locally exploiting compliance through flexible paws [Catalano21, Badri-Spröwitz22] or tails [Libby12, Liu25], has shown to increase locomotion speed and acceleration capability [Fisher17, Li23], enhance impact absorption [Sato15, Hua25], or stabilize gaits [Eckert15, Sabelhaus18]. While effective on targeted benchmarks, they remain isolated and fail to integrate into a comprehensive design. These innovations are typically investigated at the component level, without considering whole-body interactions, due to the lack of a co-design methodology that incorporates robust whole-body control policies.
This PhD project will leverage recent co-design approaches and modern control methods such as RL to go beyond component-level innovations. It aims at developing a new whole-body quadruped architecture that embodies the use of alternative mechatronic concepts, to improve overall system manipulation and dynamic locomotion capabilities, impact tolerance, and energy efficiency. It will explicitly account for control constraints through realistic simulations of the robot behavior on targeted locomanipulation tasks.
