Robust neural network controller for a lower limb exoskeleton with minimal sensor configuration: a deep reinforcement learning approach with policy distillation

Background: Lower limb rehabilitation exoskeletons (LLREs) are increasingly used in clinical settings to improve mobility in individuals with neuromuscular impairments. Most LLREs employ controllers focused on trajectory tracking, which lack adaptability to user-specific variations or voluntary effort. Autonomous LLREs enable hands-free gait but often require a multitude of sensors such as IMU, encoders, and foot force sensors to enable user intent prediction, gait phase detection, and dynamic balance. This often introduces substantial challenges related to cost, complexity, reliability, and system scalability. Methods: In this study, we introduce a novel deep reinforcement learning (DRL) based approach for autonomous control of a custom-designed LLRE using a minimal sensor configuration, enabled through a privileged teacher–student policy distillation paradigm. Policies are trained in a physics-based simulation environment integrating a full-body musculoskeletal model and an exoskeleton interaction model. The privileged teacher control policy leverages privileged full-state information to learn stable walking behaviors, while the student control policy learns to replicate the privileged control policy’s behavior via policy distillation. The student policy uses only proprioceptive signals derived from joint encoders, enabling direct deployment on physical hardware with minimal sensor requirements. Results and conclusion: We evaluate both privileged teacher and student control policies in simulated walking scenarios under external disturbances. Performance metrics such as gait symmetry and lateral stability confirm that the student policy, despite relying solely on encoder data, achieves comparable performance to the teacher policy and remains robust to disturbances. Further comparisons with sensor-rich configurations, including those incorporating IMU-based orientation and foot force sensor derived center-of-pressure (CoP), show minimal performance degradation under the joint encoder only configuration. These results highlight that robust LLRE control can be achieved with substantially reduced sensing demands. Our method supports seamless sim-to-real transfer, simplifies hardware integration, reduces calibration and fault risks, and enhances the practicality of deploying autonomous exoskeletons in both clinical and real-world environments.