Colloquia

Samenvatting

Robotic lower-limb exoskeletons offer a promising solution for paraplegic patients by enabling standing and fully supported walking, improving quality of life, and reducing health issues related to prolonged wheelchair use. To further restore the mobility and independence of users, exoskeletons should support walking through various environments and daily activities. One of the limiting factors is the generation of appropriate reference trajectories. Current methods often use parametrized trajectories of a single gait cycle, which limits gait variability, as including additional gait types increase controller complexity, storage space, and computation time. Consequently, gait transitions are often avoided by starting and ending each cycle at a stationary position. However, these interruptions are at the expense of time, energy, and user experience. Therefore, this study evaluates the feasibility of a novel trajectory generation framework designed to enable continuous walking across varying terrains with a user-selected velocity.

The proposed framework integrates concepts from computer animation with robotics by combining a Phase Functioned Neural Network (PFNN) for real-time trajectory generation with a Dynamics Filter that adjusts the trajectory to a physical feasible one for a lower-limb exoskeleton. Such a framework should eliminate the need for a complex high-level controller or a library of gait patterns and transitions. Instead, it generates whole-body reference trajectories based on user input, a trained network, a heightmap of its close environment, and a dynamic model of the exoskeleton.

Feasibility is assessed by generating trajectories for steady-state, transient, and perturbed trials. The results show that the PFNN-generated trajectories adapt to changes in velocities and inclines, following human-like trends, without interpolation artifacts. This demonstrates the multimodality and suggests intrinsically stable interpolation behaviour of the PFNN. The Dynamics Filter compensates for actuator limitations, adjusts for dynamic differences, and accounts for contact forces. Additionally, feedback of the global velocity enables the framework to respond to perturbations during quiet stance by generating step-response trajectories. While several limitations remain, including foot slips and uncaptured kinematic differences that affect the contact position and timing, the results indicate the feasibility of the proposed framework at a kinematic level.