Abstract

Prismatic-knee biped robots have recently been considered an alternative to revolute-knee biped robots. In this study, we investigate control and stabilization strategies for a three-link biped robot with point feet, actuated at the hip and knee joints, making it an underactuated system. Two architectural configurations are considered: (i) a model with a prismatic knee joint, and (ii) a model with a conventional revolute knee joint. The investigation aims to evaluate and compare how these two knee designs affect the robot’s ability to achieve dynamic stability, especially when subjected to diverse disturbances or non-equilibrium initial configurations. For both the biped robot variants, the work space of the hip joint is systematically sampled to generate a set of disparate initial conditions. These scenarios reflect practical situations where the robot may encounter substantial deviations from its standing equilibrium–such as after a push, slip, or terrain irregularity. The control policy is trained using the Soft Actor-Critic (SAC) algorithm in a custom Gymnasium environment built with the MuJoCo simulator using default frictional values. The observation space includes ground reaction forces, joint velocities, joint positions, inertia matrix, center of mass velocities.