GaitNet: Greedy, Acyclic Quadruped Gait Generation
Quadruped robots struggle to move reliably across complex, unpredictable terrain, where controllers must balance adaptability and stability. This thesis presents GaitNet, a hybrid locomotion framework that combines a neural network–based gait planner with a model-based controller to generate dynamic, non-periodic footstep patterns in real time. The system includes a footstep evaluation network and a reinforcement learning gait selector, trained in large-scale GPU simulation.
Experiments show that GaitNet more than doubles the survival rate of a strong baseline in challenging terrain and reveals key insights into the role of footstep costs and timing in multi-leg coordination. These results demonstrate that greedy, learning-based planning can produce robust, adaptable quadruped motion and offer a promising path toward fully autonomous legged locomotion.
Advisor: Professor Mahdi Agheli
Committee: Professor Jing Xiao, Professor Constantinos Chamzas
Zoom link: https://wpi.zoom.us/j/4841092774