Design and Evaluation of a Mobile Manipulation System with Learning-Based Slip Mitigation

Abstract

This work presents the design, modeling, and evaluation of a mobile manipulation system that integrates navigation, manipulation, learning-based slip estimation, and vision-based localization within a unified control framework. A key challenge addressed is wheel slip, which violates the no-slip assumptions underlying classical mobile robot kinematics and leads to degraded odometry. To address this, a supervised learning-based slip estimation model is trained using simulation-generated data. This estimator is embedded within a real-time finite-state traction controller that dynamically adapts wheel commands. In simulation, the controller reduces slip duration compared to unmitigated control; navigation, manipulation, and vision-based localization are demonstrated on a physical robot.