يادگيري گشتاور متقارن كمكي براي اسكلت‌هاي بيروني پايين‌تنه با استفاده از يادگيري تقويتي عميق

Wearable lower-limb exoskeletons traditionally face a fundamental trade-off between walking stability an‎d energy efficiency: high-gain tracking control guarantees safety at a high mechanical cost, while compliant control saves energy at the expense of robustness. This thesis develops a deep-reinforcement-learning-based assistive controller for a six-degree-of-freedom lower-limb exoskeleton, trained with Proximal Policy Optimization (PPO) in a coupled human-exoskeleton simulation. The human wearer is modelled as a periodic, self-synchronizing gait controller built on the Hybrid Zero Dynamics framework, using Bezier-parameterized virtual constraints driven by a state-based phase variable rather than time; this construction keeps the coupled system upright an‎d stable throughout training. The assistive policy is trained with a multi-objective reward that balances a saturated forward-velocity term, a per-joint-normalized energy penalty, a hard postural-stability constraint, an‎d a geometric symmetry penalty. Sample efficiency an‎d gait symmetry are further improved through mirrored-state data augmentation, an‎d adaptation to sloped terrain is achieved through a lightweight residual-dynamics architecture that leaves the flat-ground base policy untouched. At a walking speed closely matched to the unassisted baseline (about 1.00 m/s versus 1.13 m/s), the learned policy reduces the wearerʹs mechanical cost of transport by about 14% (total cost of transport of about 7.82, of which the exoskeletonʹs own share ranges from 9.5% to 14.6% across three independent training runs), while keeping torque opposition between the wearer an‎d the device low an‎d preserving upright pelvis height an‎d bilateral gait symmetry. The resulting gait recovers from external pushes of up to 150 N an‎d converges to the same efficient policy in three of five independent training runs. These findings show that pairing a virtual-constraint-based human model with a carefully designed multi-objective reward can deliver substantial energy savings in exoskeleton control without requiring new compliant hardware.

حامد جلالي بيدگلي

احسان روحاني , محمدرضا احمدزاده