Deep-Learning Control of Lower-Limb Exoskeletons via simplified Therapist Input

TL;DR

This work tackles calibration-heavy hierarchical control in partial-assistance lower-limb exoskeletons by introducing a three-step data-driven framework that (1) probabilistically maps recent sensor data to clinically relevant gait features, (2) allows therapists to adjust these features via a user interface, and (3) regresses them to a probabilistic joint posture and impedance for the device. The system employs two neural networks, a Features Extractor Model (FEM) and a Command Predictor Model (CPM), with Monte Carlo uncertainty propagation to adapt impedance parameters in real time. Validations on two healthy participants across treadmill, stairs, and ramps show the controller can infer user intent and generate mostly assistive (negative) interaction power, though some oscillations and dynamic limitations were observed. By leveraging uncertainty to modulate stiffness and damping, the approach reduces calibration burden and enables therapist-driven personalization, with potential clinical impact for gait rehabilitation.

Abstract

Partial-assistance exoskeletons hold significant potential for gait rehabilitation by promoting active participation during (re)learning of normative walking patterns. Typically, the control of interaction torques in partial-assistance exoskeletons relies on a hierarchical control structure. These approaches require extensive calibration due to the complexity of the controller and user-specific parameter tuning, especially for activities like stair or ramp navigation. To address the limitations of hierarchical control in exoskeletons, this work proposes a three-step, data-driven approach: (1) using recent sensor data to probabilistically infer locomotion states (landing step length, landing step height, walking velocity, step clearance, gait phase), (2) allowing therapists to modify these features via a user interface, and (3) using the adjusted locomotion features to predict the desired joint posture and model stiffness in a spring-damper system based on prediction uncertainty. We evaluated the proposed approach with two healthy participants engaging in treadmill walking and stair ascent and descent at varying speeds, with and without external modification of the gait features through a user interface. Results showed a variation in kinematics according to the gait characteristics and a negative interaction power suggesting exoskeleton assistance across the different conditions.

Figures (5)

• Figure 1: (1) Data-driven controller:(1.A) Sensor readings are passed to multiple deep-learning models to estimate representative features of the walking pattern. Thes resulting features are fitted to a normal distribution to consider the uncertanty of the locomotion pattern; (1.B) User interface allows to modify the self-selected locomotion features to allow the therapist to adapt the locomotion depending on the needs of each patient; (1.C) The resulting modified features are passed to a combination of models to regress the desired joint configuration of the exoskeleton. At the same time, the uncertainty of the prediction is propagated and used to model the stiffness of the impedance controller of the exoskeleton. Impedance parameters are passed to the low-level controller of the exoskeleton as commanded input. (2) Structure of the feature extractors models. (3) Structure of the command predictor models.
• Figure 2: Example of walking pattern for a single subject. In order, the five rows show: the implemented features (self-selected or operator-selected) over time [s], the reference position for the left hip and left knee, and the reference velocities for left hip, and left knee over the left stride percentage [%]. During the trial the user performed: (A) walking with exoskeleton controlled in transparency (self-selected features are not used to close the control loop) with treadmill velocity equal to 0.14 m/s; (B) Self-selected features used to close the loop of the controller; (C) walking with self-selected features on the treadmill at 0.19 m/s; (D) Operator-selected step clearance; (E) Operator-selected step velocity (at the same time increasing the treadmill velocity to 0.19 m/s); (F) Operator-selected step length.
• Figure 3: Example of gait feature prediction. The plot displays the ground truth (in green) and the prediction of the Features Extractor Model (in blue) for two users in the training dataset over time (in seconds). Each user (user 1 in pink, user 2 in yellow) performed in order overground walking, ramps (both ascending and descending), and stairs(both ascending and descending). In order top-down the following features are displayed: step height, step length, step velocity, and step clearance.
• Figure 4: Joint kinematic across gait features (both self-selected and operator-selected): The two rows display respectively the hip and knee joint angles (left and right grouped). Each column represents with a different color code the distribution of the joint kinematic for each gait feature (step-height, step-length, step-velocity, step-clearance). On the top of each column are displayed the color bars for each gait feature. The trend was grouped based on similarities in gait features and is presented using the mean and standard deviation.
• Figure 5: Joint power across gait features (both self-selected and operator-selected): The two rows display the hip and knee joint power (left and right grouped). Each column display with a different color code the distribution of the desired joint kinematic for each gait feature (step-height, step-length, step-velocity, step-clearance).On the top of each column are displayed the color bars for each gait feature. The trend was grouped based on similarities in gait features and is presented using the mean and standard deviation. Negative joint power means exoskeleton assistance to the user while positive joint power represents resistance to the user.