Virtual environments provide a safe and accessible way to test innovative technologies for controlling wearable robotic devices. However, to simulate devices that support walking, such as powered prosthetic legs, it is not enough to model the hardware without its user. Predictive locomotion synthesizers can generate the movements of a virtual user, with whom the simulated device can be trained or evaluated. We implemented a Deep Reinforcement Learning based motion controller in the MuJoCo physics engine, where autonomy over the humanoid model was shared between the simulated user and the control policy of an active prosthesis. Despite not optimising the controller to match experimental dynamics, realistic torque profiles and ground reaction force curves were produced by the agent. A data-driven and continuous representation of user intent was used to simulate a Human Machine Interface that controlled a transtibial prosthesis in a non-steady state walking setting. The continuous intent representation was shown to mitigate the need for compensatory gait patterns from their virtual users, and halve the rate of tripping. Co-adaptation was identified as a potential challenge for training human-in-the-loop prosthesis control policies. The proposed framework outlines a way to explore the complex design space of robot-assisted gait, promoting the transfer of the next generation of intent driven controllers from the lab to real-life scenarios.

For the control of wearable robotics, it is critical to obtain a prediction of the user’s motion intent with high accuracy. Electromyography (EMG) recordings have often been used as inputs for these devices, however bipolar EMG electrodes are highly sensitive to their location. Positional shifts of electrodes after training gait prediction models can therefore result in severe performance degradation. This study uses high-density EMG electrodes to simulate various bipolar electrode signals from four leg muscles during steady-state walking. The bipolar signals were ranked based on the consistency of the corresponding EMG envelope’s activity and timing across gait cycles. The locations were then compared by evaluating the performance of an offline Temporal Convolutional Network (TCN) that mapped EMG signals to knee angles. The results showed that electrode locations with consistent EMG envelopes resulted in greater prediction accuracy compared to hand-aligned placements (p<0.01). However, performance gains through this process were limited, and did not resolve the position shift issue. Instead of training a model for a single location, we showed that randomly sampling bipolar combinations across the high-density EMG grid during training mitigated this effect. Models trained with this method generalised over all positions, and achieved 70% less prediction error than location specific models over the entire area of the grid. Therefore, the use of high-density EMG grids to build training datasets could enable the development of models robust to spatial variations, and reduce the impact of muscle-specific electrode placement on accuracy.