Learning-Augmented Motion Planning and Control for Legged and Hybrid Locomotion

Replicating physical intelligence—the ability of natural systems to perceive, reason, act, and learn in the physical world—has long been a central ambition of robotics, and remains far from being achieved. Within this broad objective, this thesis focuses on artificial motor intelligence for legged and hybrid robotic systems: the ability to plan and control motion under dynamics, contact interactions, actuation limits, and uncertainty. This thesis argues that progress in contact-rich motor intelligence does not require abandoning models in favor of data-driven methods. Rather, models provide the structural bias needed for sample-efficient planning, interpretable and safer control, and robust online decision making under constraints, while learning becomes valuable when optimization alone becomes too expensive or when key aspects of the task, such as contact scheduling, are difficult to encode explicitly. The first contribution investigates physics-rich offline motion planning for agile jumping with an electrically actuated leg. A two-stage trajectory optimization pipeline is developed to optimize the take-off and landing-braking phases. The proposed formulation explicitly optimizes take-off timing, while also accounting for actuator behavior, impacts, energy regeneration, and an identified friction model regressed from data, aspects that are often neglected in motion planning. Although limited to open-loop motion generation and to settings with very few contacts, this study shows that increasing model completeness can substantially improve planning accuracy and the efficiency with which the system exploits its own dynamics. The second and central contribution of the thesis is a learning-augmented MPC architecture for non-gaited legged and hybrid locomotion. Moving beyond the previous single-contact setting and addressing the fixed-gait assumption of traditional locomotion MPC, the proposed framework augments a full rigid-body MPC with a learned RL agent for contact scheduling and navigation. The resulting architecture combines efficient Differential Dynamic Programming (DDP)-based MPC, full rigid-body dynamics models, and data-efficient off-policy Soft Actor-Critic (SAC) learning. By exploiting low-cost online policy evaluation, it retains real-time performance across platforms with different morphologies and weight distributions, while enabling zero-shot transfer across simulation domains and to hardware without domain randomization. This line of work is supported by the development of IBRIDO, a modular and open-source software framework for learning-augmented MPC. Natural extensions include learning more general policies that transfer across controllers and platforms, and enriching the action and observation spaces for terrains and tasks beyond the non-flat scenarios already considered. The third and final contribution addresses the same contact-scheduling problem from a complementary direction by developing differentiable contact dynamics and sensitivity regularization for contact-implicit optimization. The main premise is that differentiability alone is not sufficient in contact-rich systems: gradients through contact transitions must also remain informative enough to support optimization. The proposed formulation evaluates candidate trajectories with a nonsmooth rigid-contact simulator, while using relaxed backward sensitivities to provide useful search directions through contact creation, contact loss, and frictional transitions. Current outcomes include extensive testing of the contact dynamics, derivative verification, and offline contact-implicit trajectory-optimization results showing that relaxed sensitivities can improve optimization through contact reconfiguration. This contribution therefore opens a path toward CI-MPC, learning-augmented CI-MPC and, ultimately, contact-implicit adaptive MPC.