Torque-Level NMPC for Robust End-Effector Control Under Quadruped Base-Induced Disturbances

Quadrupedal robots' ability to perform precise manipulation tasks while maintaining mobility on uneven terrain has continued to be an area of interest. However, achieving such capability is challenging, as stable manipulation can become difficult due to disturbances introduced by the robot’s gait on uneven surfaces. Inspired by "chicken head" stabilization, this study aims to investigate the effectiveness of Nonlinear Model Predictive Control (NMPC) in maintaining a stable end-effector of a robotic arm mounted on a quadruped robot. To achieve this goal, a ROS 2/Gazebo simulation environment was developed, with NMPC optimization implemented through CasADi and acados to compute joint torques in real-time. The simulation introduces disturbances through uneven terrain and deliberate base movements that represent different gait patterns. Results demonstrate that the controller regulates the end-effector toward the desired pose under both stationary (1.3 cm mean error) and dynamically-disturbed (1.7 - 3.6 cm mean error) base conditions, with residual oscillations attributed to simulation artifacts and controller sensitivity. These results demonstrate that decoupled torque-level NMPC can achieve end-effector regulation under base disturbances and establishes a framework for general mobile manipulation, treating the base as a generic disturbance source, with potential applications in planetary exploration, maritime, and in-orbit servicing.