Model-Based Compensation for Serial Manipulators through Semi-Parametric Gaussian Process Regression

Abstract

Industrial robots can be found in automotive, food, chemical, and electronics industries. These robots are often caged and are secluded from human beings. A recent trend in a subclass of industrial robots named collaborative robots allows the humans to interact with the robots safely. The word “safety” mentioned above is of supreme importance. The safety is achieved in these robots by their lightweight and sleek design. Often, robots are operated under low stiffness conditions to achieve less impact force during an unavoidable collision. A severe damage to the environment may occur if the robot becomes unstable under any conditions. It is of paramount importance for the controller present in the robot to stabilize the system under all conditions. One such controller is the joint impedance controller, which helps the robot to interact with an unknown environment by causing no harm to humans.

The thesis marks its importance, as it is closely related to ensuring safety in collaborative robots and is mainly focused on tackling the situations whenever the controller fails. The controller in these manipulators has an Inverse Dynamics Model (IDM) and a Proportional Derivative (PD) controller. Under low stiffness and damping conditions, the PD gains are low and the manipulator is entirely compensated by the inverse dynamics model. This inverse dynamics model can become problematic in the presence of the un-modeled dynamics like flexibility, friction, dynamics of hydraulic tubes, actuators and cable drives or if the IDM model is inherently inaccurate. Consequently, the in-built joint impedance and position controller will fail to work under low stiffness and damping conditions, in-turn making the robot unstable. If this robot was to be used on an industrial platform and the problem is unresolved, it might cause some danger to the humans working closely and also damage the environment
and itself.

Since the robot is entirely compensated by the IDM under low stiffness and damping conditions, the thesis tries to acquire the accurate IDM of the robot for control purposes. To do so, two cases were modeled in this thesis, one with the internal IDM with correct base parameters and another one with the incorrect internal IDM by adding offset in the base parameters. But in both cases, the internal IDM model failed to compensate for the un-modeled dynamics
occurring in the manipulator.

The thesis incorporates a semi-parametric Gaussian process regression to tackle the two cases. A semi-parametric model consists of a parametric term and a non-parametric term. First, the parametric term is identified using the least squares approach. Later, the parametric term is used as mean to capture the non-parametric term using the Gaussian Process Regression. The proposed methods were tested on the PUMA 560 robot and the two-link manipulator in MATLAB. From the simulation results, the semi-parametric model was able to provide accurate feed-forward control torques to compensate for the model inaccuracies and the un-modeled dynamics at low stiffness and damping conditions. Additionally, implementing these proposed methods on a real robot will be a future scope of improvement on this topic.