Text

Introduction: Occupational exoskeletons are a promising solution to physically releive workers in strenuous tasks, and reduce the prevalence of musculoskeletal disorders (De Looze M. Ergonomics, 2016). Active exoskeletons offer more versatility than passive devices, since the assistance can be adjusted to the context, the human motor intention or physiological state. However, the design of efficient control laws usually requires to collect human-exoskeleton motion data, for tuning parameters of the control law, or as training data for machine learning based control approaches. But such data collection is costly and lengthly, and thus an obstacle to exoskeleton design.

Methods: We propose to leverage physics-based simulation of the human-exoskeleton system to generate synthetic data of human-exoskeleton interaction, that can be used for pre-training or pre-tuning exoskeleton control. This simulation allows us to estimate biomechanical quantities, e.g. human posture and internal forces. We use quadratic programming, a control method stemming from humanoid robotics, to generate the motion of the human-exoskeleton system. In a previous work, we proposed to track motion capture data collected without exoskeleton with a digital exoskeleton, in order to emulate the sensor data of the exoskeleton. We used these data to train predictive models of human movements, and showed that adding this prediction to a standard exoskeleton control law improves the system transparency.

Here, we extend the previous approach by adding the physical coupling between the human and the exoskeleton in the simulation. We then use the simulation in an optimization process to tune the parameters of the predictive control law.

Results: We tested our approach on a repetitive pointing task involving load manipulation, assisted by an elbow exoskeleton.

Estimating realistic values for the parameters of the human-exoskeleton contact model is difficult. Hence, we conducted a sensitivity analysis to evaluate how much these parameters affect the simulated human data. Our results show that while the human-exoskeleton interaction force is logically very sensitive, the simulated human motion and forces are more robust, and thus can be reliably used.

We then optimized two parameters of the predictive control law using Bayesian optimization. The objective was to minimize the elbow joint torque difference with respect to the no-load no-assistance motion. The parameters optimized in simulation were then successfully used in a real subject experiment with one pilot subject. The subjective feedback of the participant suggests that the simulation-based tuning can be used as an efficient initial guess.

Discussion: The proposed simulation-based approach to generate synthetic data of human-robot interaction showed promising results. The quadratic programming based approach enables real-time simulation and can account for the effect of the assistance when generating the human motion. We are currently validating the approach with more participants. Future work includes developing models of human motor adaptation to the assistance to enhance the realism of the simulated assisted motion.