Towards Optimal Gait Generation: A Comparative Study of Optimization-Based, Classical (LIPM-ZMP), and Reinforcement Learning Approaches for NU-Biped-V4.7 robot

Abstract

This thesis examines one of the central topics in robotics: bipedal locomotion. While there are numerous approaches to gait generation, there is no universal solution, making the portability of methods across platforms a complex, time-consuming, and sometimes unfeasible process, especially for custom robots. The goal of this study is to implement and compare different methods for generating gait on the NU-Biped-V4.7 robot in the PyBullet simulation. To address this problem, three approaches were considered in this work: optimization-based transfer of a reference gait, the classical method based on LIPM and ZMP, and a reinforcement learning method.To provide a fair qualitative comparative evaluation, all methods were tested under identical environmental conditions, and their performance was assessed in terms of stability, smoothness, and feasibility. The results show that classical and optimization methods allow for a stable gait but require careful tuning, whereas reinforcement learning methods are more adaptive but face convergence issues. The work demonstrates that there is no universal approach, and the choice of method is determined by the specific characteristics of the robotic system.