Sparse Online Gaussian Process Impedance Learning for Multi-DoF Robotic Arms

Robot interaction control with fixed impedance parameters usually fails to achieve the desired interaction behavior, motivating the introduction of variable impedance or admittance control. A gradient-descent (GD) impedance learning approach with a limited iteration number can make robots more compliant by minimizing an objective function. However, due to the nature of GD optimization, impedance parameters can not be adjusted in time by this approach, resulting in degraded robot compliance. This paper combines GD optimization with sparse online Gaussian process (SOGP) to develop a GD-based SOGP (GD-SOGP) impedance learning approach for variable admittance control. A high-fidelity mathematical model of a 7-DoF collaborative robot called Panda is applied for simulation studies. It is shown that the proposed GD-SOGP impedance learning can make the robot more compliant and outperforms the GD impedance learning in terms of impedance convergence.