The defining characteristic of humanoid robots is their human-like bipedal locomotion, which enables autonomous movement and interaction within human environments. For this purpose, convex Model Predictive Control (MPC) based on Single Rigid Body Dyna...
The defining characteristic of humanoid robots is their human-like bipedal locomotion, which enables autonomous movement and interaction within human environments. For this purpose, convex Model Predictive Control (MPC) based on Single Rigid Body Dynamics (SRBD) has been widely adopted due to its high computational efficiency. However, by approximating the robot as a single rigid body, SRBD introduces significant modeling errors in bipedal systems where the legs contribute a large portion of the total mass and inertia. This thesis aims to design a walking control framework for a humanoid robot with 5-DoF legs that preserves the real-time advantages of SRBD MPC while compensating for these modeling limitations.
To address this issue, two control components are added to the SRBD MPC framework. A Whole-Body Control (WBC) layer is introduced downstream of the MPC to generate whole-body torque commands using full-body dynamics and null-space–based hierarchical control. In addition, an Inertia-Adaptive MPC (IA MPC) replaces the constant centroidal inertia tensor in the SRBD model with time-varying inertias predicted by a Centroidal Composite Inertia Neural Network (CCINN), which are applied at each step of the MPC prediction horizon while preserving convexity and fast computation.
To evaluate the performance of the developed control framework, experiments were conducted in both Gazebo simulation and real hardware. In simulation, the robot fell before reaching 10 m when using only SRBD MPC, while SRBD MPC + WBC enabled a 50 m walk with a lateral deviation of 10.9 m. In contrast, the IA MPC + WBC controller reduced the deviation to 0.51 m and decreased Roll RMSE by more than 50%. On real hardware, the IA MPC + WBC controller achieved stable straight-line walking and reduced Roll orientation error by more than 70% compared to SRBD MPC + WBC. These results demonstrate that integrating whole-body control with adaptive centroidal inertia modeling significantly enhances bipedal walking stability.