Fig. 3: Assist action sequence from our approach and performance against EMG-based methods. | npj Robotics

Fig. 3: Assist action sequence from our approach and performance against EMG-based methods.

From: Transformer-based multitask assist control from first-person view image and user’s kinematic information for exoskeleton robots

Fig. 3

a The red solid line shows the predicted sequence, and the blue dashed line shows the actual action sequence annotated by the user for each leg motion. Act and Fr of the pressure command indicate “active" and “free," respectively. The corresponding torque to the pressure command is calculated based on Eqs. (1) and (2). b F-score of the average of right and left leg motions when the estimations were treated as a binary classification. VK-TR refers to our approach. EK-T uses the same transformer architecture as the proposed approach with EMG and kinematic information, but does not use ResNet. EK-S and EK-LD use a Support Vector Machine with a radial basis function kernel and Linear Discriminant Analysis with EMG and kinematic information, respectively. Our proposed method was able to achieve comparable or better estimation performance with EMG-based methods, even without using any bio-signals. Our method does not need careful preparation and calibration, where these cumbersome procedures prevent us from adopting bio-signals for assistive robot control for daily use.

Back to article page