Abstract
Human gait prediction enables anticipatory assistance in hip exoskeletons, where safe and timely support requires adjusting joint behavior before the next step rather than reacting to deviations that have already occurred. This thesis examines whether three clinically relevant step-level parameters—step length ˆ L, mid-swing foot height ˆ H, and mid-swing velocity ˆV—can be forecast one step ahead using reduced inertial measurement unit (IMU) configurations compatible with practical exoskeleton sensing. Together, these descriptors provide a compact representation of gait dynamics suitable for identifying potentially destabilizing steps in advance. A further aim is to integrate these predictions into a stability-aware control model capable of generating corrective hip torques before instability arises. To address this aim, two datasets were collected. The first is a multimodal dataset combining IMUs, motion capture, depth imaging, and eye tracking across straight, curved, and obstacle-circumnavigation walking. The second is an IMU-only dataset tailored to exoskeleton-compatible sensor placement and focused on curved walking, where natural asymmetries and increased mediolateral demands create a challenging prediction context. More than 38,000 labeled steps from linear, non-linear, and cluttered environments were used to train long short-term memory (LSTM) models based on short temporal windows of IMU signals. These achieved high accuracy in straight walking, while performance declined moderately during turning yet remained robust. For obstacle negotiation, an LSTM–multilayer perceptron (MLP) fusion model encoding obstacle geometry improved height predictions in situations requiring increased clearance, although IMU-only models were more stable across layouts. Sensor placement analysis further showed that hip- and thigh-mounted IMUs—or even a single pelvis unit—maintained performance within approximately 2% of full lower-limb setups. Building on these findings, the thesis introduces a predictive stability-aware control framework that maps next-step predictions into continuous hip trajectories for exoskeleton control. Predicted and geometry-scaled corrected parameters define stance and swing boundary conditions, ensuring motion remains within normative safety limits. Using an inverted-pendulum representation, center-of-mass, extrapolated center-of-mass, and margin-of-stability profiles are forecast for the upcoming step, enabling margin of stability (MoS)-dependent modulation of hip impedance. Overall, the thesis demonstrates that accurate one-step-ahead prediction under reduced sensing is feasible across diverse locomotor tasks and can be effectively integrated into a stability-oriented control framework for hip exoskeletons. The work advances dataset design, reduced-sensor gait modelling, and environment-aware prediction, supporting the development of exoskeletons capable of proactive rather than reactive intervention.