A research article published by the Peking University presented a control framework for exoskeletons based on environmental perception, which effectively integrates environmental information and human kinematic data, improves the accuracy and lead time of transition detection, thereby enhancing smooth switching of control strategies across different terrains. Additionally, the adoption of a learning-free method eliminates the need for data collection and model training, demonstrating strong generalization capabilities across users.
The new research paper, published on Apr. 22 in the journal Cyborg and Bionic Systems, presented a study on adaptive control of hip exoskeletons in level-ground and stair environments employs a continuous locomotion mode perception technology based on a learning-free (non-data-driven) method.
Recent advances in hip exoskeleton control have demonstrated potential in enhancing human mobility across diverse terrains. However, achieving seamless adaptation to continuous locomotion modes (e.g., level-ground walking, stair ascent/descent) without user-specific data training remains a significant challenge. "By integrating depth-enhanced visual-inertial odometry and terrain reconstruction, our learning-free method eliminates dependency on datasets while maintaining high prediction accuracy across subjects," explained corresponding author Qining Wang, a professor at Peking University. The proposed three-layer control framework incorporates (a) a depth camera for real-time environment mapping, (b) pressure insoles for gait phase detection, and (c) physics-driven torque/damping strategies tailored to biomechanical profiles. "This approach ensures smooth transitions between control modes by predicting terrain changes before the end of transition periods," added lead author Zhaoyang Wang.
The system was validated through experiments with 7 subjects performing continuous locomotion (LG, SA, SD, and transitions). High-level perception achieved >95% accuracy for steady modes (LG: 98.1% ± 1.8%, SA: 97.3% ± 3.7%, SD: 95.8% ± 3.9%) and 87.5–100% accuracy for transitions, detecting shifts 14.5–30.5% earlier than transition completion. Mid-level control employed phase-specific torque curves (e.g., peak extension torque at 42% gait cycle for LG) and constant damping for SD, while low-level execution utilized PID-based torque tracking and PWM-modulated braking. Compared to CNN-based methods, this framework improved transition accuracy by 20–30% and reduced reliance on user calibration.
"Vision-based terrain reconstruction enabled precise prediction of stair geometry, aligning exoskeleton assistance with real-time biomechanical demands," noted Wang. Limitations include sensitivity to lighting variations and unstructured environments. Future work will integrate multimodal sensors and validate efficacy in clinical populations. This study advances adaptive exoskeleton control by merging environmental intelligence with human-centric design, paving the way for robust, user-agnostic assistive technologies.
Authors of the paper include Zhaoyang Wang, Dongfang Xu, Shunyi Zhao, Zehuan Yu, Yan Huang, Lecheng Ruan, Zhihao Zhou, and Qining Wang.
This work is supported in part by the National Natural Science Foundation of China (nos. 91948302, 52005011, 62073038, and 52475001) and Beijing Municipal Science and Technology project (no. D181100000318002).
The paper, "Level-Ground and Stair Adaptation for Hip Exoskeletons Based on Continuous Locomotion Mode Perception" was published in the journal Cyborg and Bionic Systems on Apr 22 2025, at DOI: 10.34133/cbsystems.0248.
