Control of robots with hybrid locomotion capabilities

Hybrid locomotion robots—systems combining multiple modes of motion—are well-suited for challenging terrain and have broad practical applications. However, developing effective control strategies remains challenging due to nonlinear dynamics, multimodal locomotion, computational constraints, and multiple objectives. Existing solutions are often not portable to unconventional morphologies, requiring substantial redesign. This thesis proposes multiple control solutions for three morphologically distinct robots: Asguard, SherpaTT, and ARTER.

Asguard’s five-spike wheel design presents challenges in forward motion and point turning. A cascaded position–velocity–torque controller improves wheel positioning accuracy by up to 56 % while maintaining defined offsets. The controller based on a novel torque estimator using mechanical coupling deflection enables real-time torque control. Adjusting inter-wheel offsets reduces resistance to motion by up to 90 %. For point turns on rough terrain, an algorithm that utilizes external load torques enables reliable rotation, outperforming baseline controllers.

Dedicated control frameworks, denoted as Motion Control System (MCS), are developed for the wheel-legged robots SherpaTT and ARTER. SherpaTT-MCS supports teleoperation, assistance functions, and autonomy. Its terrain adaptation module improves force distribution by 80 % and reduces attitude error by up to 95 % in laboratory tests. It has been successfully deployed in over ten research projects and has proven its efficacy in three Mars-analogous field trials. ARTER-MCS incorporates kinematic modeling of parallel linkages and nonlinear model predictive control, and is currently in active use in multiple projects.

For ARTER, a Deep Reinforcement Learning (DRL)-based terrain adaptation controller is introduced, leveraging compressed height-maps via autoencoders. Ten variants of the controller were trained using different combinations of observations, including contact distances and latent-space representations. The controller that demonstrated the strongest performance utilized contact-detection and a 4-dimensional terrain latent space as observation, offering a favorable combination of both performance and complexity. All controllers achieved baseline objectives and has the potential to be ported to other platforms with active suspension.

ARTER also demonstrates stepping locomotion via a controller that combines the movement of the manipulator arm, the legs and the wheels. This controller applies hierarchical reinforcement learning and action masking, integrating domain knowledge to simplify training. A three-level hierarchy is employed: the lowest level manages diverse simpler motions (manipulator, longitudinal motion, end-effector height adjustments, etc.); the middle level sequences these for stepping in and out of obstacles; the top level manages task transitions. The architecture generalizes across three different types of stepping terrain and generated motion comparable to that of an expert operator.

In summary, this thesis presents a series of control solutions designed to enhance the locomotion performance, efficiency, and adaptability of hybrid robotic platforms. The learning-based methods offer strong morphological generalization and address long-sequence tasks with reduced engineering effort. These contributions represent quantitative and qualitative advancements in the control of diverse robots with hybrid locomotion capabilities.