Embedded Adaptive Admittance Control for a Modular Hip Exoskeleton Using Deep Learning-Based Gait-Phase Estimation

Hip assistance robots have become a practical solution as daily-life walking support. However, providing reliable and timely assistance during overground walking remains challenging because gait varies across users and walking conditions. This study presents HealBot-H, a newly developed 2-kg modular hip exoskeleton with a gait-phase-aware admittance control framework driven by a deep learning-based gait phase estimator. The robot features two active hip joints in the sagittal plane and uses a magnetic quick-release modular structure, enabling either unilateral or bilateral configuration depending on the application. The control framework combines a classical admittance law with a bidirectional long short-term memory network that estimates locomotion mode and the continuous gait phase from lower-limb inertial measurement units. The trained model was deployed on a Raspberry Pi 5 for real-time operation. Based on the estimated gait phase, the admittance parameters are scheduled across seven subphases of the gait cycle. To avoid abrupt torque changes at the subphase boundaries, a two-stage update rule comprising linear interpolation and exponential smoothing is applied. The robot was validated with ten healthy adults during overground walking. Surface electromyography (EMG) was recorded from five lower-limb muscles and compared between walking with and without the robot. The mean EMG reductions ranged from 19.0–23.9% over the gait cycle after false discovery rate correction (all pFDR<0.01). These results demonstrate that gait-phase-aware admittance control on a lightweight wearable platform can provide effective and well-timed walking assistance, while establishing a foundation for future personalized assistance.