Legged Robot Running Using a Data-Driven Neural-Network Motion Template

We report on developing a data-driven methodology to model the robot’s dynamic behavior and then utilize this model to further generate a new robot’s dynamic behavior with improved performance and a closer match to the model’s prediction. First, the neural network (NN) architectures were evaluated using the second-order dynamical systems with given analytical solutions. The best-performing LSTM NN model was then utilized to model the dynamic running behavior of the RHex-style robot, following standard training and testing procedures. The experimental results reveal that the NN model provides better accuracy in predicting the dynamic motion of the robot than the physics-based R-SLIP model and the R-SLIP-GP model with added compensation using the Gaussian Process. Subsequently, the trained NN model served as the motion template to initiate the new dynamic running of the robot as the anchor by using the fixed-point motion of the model as the motion reference. The experimental results demonstrate that the robot exhibits more stable vertical dynamics, maintaining similar vertical behavior while experiencing smaller ground reaction forces. These results further suggest that the NN model effectively improves energy conversion and reduces the magnitude of force interactions between the robot and the ground.