Joint Moment Estimation for Hip Exoskeleton Control: A Generalized Moment Feature Generation Method

TL;DR

The paper tackles the personalization gap in hip joint moment estimation for exoskeleton control by introducing generalized moment features (GMF), generated from body parameters and ground-truth moments to yield a subject-independent representation. A GRU-based estimator learns to predict GMF from kinematic inputs, while a decoder translates GMF back to joint moments for real-time torque control. Joint training with dual losses ensures GMF is both predictive and decodable, enabling accurate cross-user predictions and efficient deployment on encoder-only hip exoskeletons. Experimental results show improved RMSE ($\approx0.1180$ Nm/kg) across 28 speeds, strong generalization, and notable metabolic cost reductions (up to $20.5\%$) during level-ground walking, confirming the approach's practical value.

Abstract

Hip joint moments during walking are the key foundation for hip exoskeleton assistance control. Most recent studies have shown estimating hip joint moments instantaneously offers a lot of advantages compared to generating assistive torque profiles based on gait estimation, such as simple sensor requirements and adaptability to variable walking speeds. However, existing joint moment estimation methods still suffer from a lack of personalization, leading to estimation accuracy degradation for new users. To address the challenges, this paper proposes a hip joint moment estimation method based on generalized moment features (GMF). A GMF generator is constructed to learn GMF of the joint moment which is invariant to individual variations while remaining decodable into joint moments through a dedicated decoder. Utilizing this well-featured representation, a GRU-based neural network is used to predict GMF with joint kinematics data, which can easily be acquired by hip exoskeleton encoders. The proposed estimation method achieves a root mean square error of 0.1180 Nm/kg under 28 walking speed conditions on a treadmill dataset, improved by 6.5% compared to the model without body parameter fusion, and by 8.3% for the conventional fusion model with body parameter. Furthermore, the proposed method was employed on a hip exoskeleton with only encoder sensors and achieved an average 20.5% metabolic reduction (p<0.01) for users compared to assist-off condition in level-ground walking.

Figures (10)

• Figure 1: Joint moment estimation for hip exoskeleton assistance based on GMF generation. During the training process, kinematics information ($\boldsymbol{X}$) is utilized as input for the estimator, while body parameters ($\boldsymbol{q}$) and hip joint moments ($M$) serve as inputs for the generator. The decoder takes body parameters and the GMF generated by the generator (represented by "*") as inputs. The estimator and decoder are then employed to estimate hip joint moments ($\hat{M}$) to provide assistance torque during walking for new users (the generator is not needed).
• Figure 2: The fast estimator based on GRU, the computation process of one output within the GRU unit, denoted as $h^0$, is shown in the solid box. The computation process for the remaining 15 hidden layer nodes, inputs ($h_{t-1}^1$ - $h_{t-1}^{15}$), and outputs ($h_t^1$ - $h_t^{15}$), follows the same process.
• Figure 3: Exoskeleton and controller used for validating the proposed method. Motors with encoder provide the controller hip joint angles as inputs and generate assistance torque to assist users. The controller is a three-level structure, where the high-level estimates hip joint moments of the user, the mid-level processes hip joint moments to obtain the desired assistance torque and the low-level FOC generates motor torque stably.
• Figure 4: Visualization of the input features distribution (joint kinematics) of the training set (in blue) and the test set data (in red) using the T-SNE method. The varying shades of color represent data from 10 individual subjects in each set. Under the data partitioning method that maximizes individual differences, the intra-subject distribution of input features within the same set (intra-class distance) appears to be similar, while the distribution of input features between different sets (inter-class distance) exhibits significant differences.
• Figure 5: Four joint moment estimation models consisting of estimators with different neural network structures, including the proposed GRU-based method for fast estimation and three mainstream model architectures.