A Learning-Based Adaptive Compliance Control Approach with Coupling Dynamics for Mobile Manipulators

Compliant force control for mobile manipulators is challenging, mainly due to two factors: (1) the dynamic coupling between the subsystems of the manipulator and the mobile base; (2) the uncertainty in the interaction between the end-effector and the environment surface. To address these issues, we propose a learning-based adaptive compliance control framework which consists of a coupling dynamics compensation loop and a DRL impedance learning loop. Based on the Koopman operator, the former captures the unmodeled coupling term in real time through online fine-tuning. The latter pretrains the feedforward neural network in the simulation and the converged policy is used to correct the impedance parameters online. To ensure the adaptability of the policy, we randomize the physical parameters and geometry of the surface the end-effector needs to interact with. A surface cleaning experiment is conducted, and the results verify the effectiveness of our proposed control scheme.