Fig. 7: Architecture of assist action prediction. | npj Robotics

Fig. 7: Architecture of assist action prediction.

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

Fig. 7

Our approach to generating assist action sequences mainly uses ResNet18 and transformer models. This model predicts sequences a few milliseconds ahead of sensor inputs that contain historical information from a few milliseconds before the current time. Predicted assist action sequence is averaged by the temporal ensemble and used for the final control command.

Back to article page