Efficient Force Control Learning System for Industrial Robots Based on Variable Impedance Control

Abstract

Learning variable impedance control is a powerful method to improve the performance of force control. However, current methods typically require too many interactions to achieve good performance. Data-inefficiency has limited these methods to learn force-sensitive tasks in real systems. In order to improve the sampling efficiency and decrease the required interactions during the learning process, this paper develops a data-efficient learning variable impedance control method that enables the industrial robots automatically learn to control the contact force in the unstructured environment. To this end, a Gaussian process model is learned as a faithful proxy of the system, which is then used to predict long-term state evolution for internal simulation, allowing for efficient strategy updates. The effects of model bias are reduced effectively by incorporating model uncertainty into long-term planning. Then the impedance profiles are regulated online according to the learned humanlike impedance strategy. In this way, the flexibility and adaptivity of the system could be enhanced. Both simulated and experimental tests have been performed on an industrial manipulator to verify the performance of the proposed method.

Keywords: Gaussian processes; efficient learning; force control; industrial robot; variable impedance control.

Figure 1

The interaction model of the system. (a) Without any contact between the robot and the environment; (b) critical point when contact occurs; (c) stable contact with the environment; (d) contact force diagram when robot comes in contact with the environment.

Figure 2

Contact force and moment applied at the end-effector.

Figure 3

An implementation example of the contact force observer.

Figure 4

The position-based impedance control schematic.

Figure 5

Scheme of the data-efficient learning variable impedance control.

Figure 6

A conceptual illustration of long-term predictions of state evolution.

Figure 7

The block diagram of simulation in MATLAB Simulink.

Figure 8

Simulation results of force control learning system. (a) The cost curve of learning process; the blue dotted line and the blue shade are the predicted cost mean and the 95% confidence interval. (b) The performances of force control during learning process.

Figure 9

Learning process of the total 20 learning iterations. The first row: contact force in z-axis direction; the second row: Cartesian stiffness schedules; the third row: Cartesian damping schedules; the fourth row: position of the end-effector in z-axis direction; the fifth row: velocity of the end-effector in z-axis direction.

Figure 10

Joint trajectories after 4, 5, 10, 15, and 20 updates for the second, third, and fifth joint of the Reinovo robot.

Figure 11

States evolution of force control. Columns (a–d) are the state evolutions of the 1st, 6th, 12th, and 20th learning iteration, respectively. The top row is the change of contact force $F_z$, the second row is the target stiffness $K_{dz}$, and the third row is the target damping $B_{dz}$.

Figure 12

Hardware architecture of the system.

Figure 13

Implementation diagram of the algorithm.

Figure 14

(a) Experimental setup; (b) simplified model of the contact environment.

Figure 15

Experimental results of force control learning system. (a) The cost curve of learning process. (b) The performances of force control during learning process, including a total 20 learning iterations throughout the experiment.

Figure 16

Main iterations of the learning process. (a) Contact force; (b) Cartesian stiffness schedules; (c) Cartesian damping schedules.

Figure 17

States evolution of force control. Columns (a–d) are the state evolutions of the 1st, 6th, 12th, and 20th learning iteration, respectively. The top row shows the contact force $F_z$, the second row shows the profile of stiffness $K_{dz}$, and the third row shows the profile of damping $B_{dz}$.

Figure 18

Trajectories during the 20th experiment iteration. (a) Joint position; (b) Joint velocity; (c) Cartesian position of the end-effector; (d) Cartesian velocity of the end-effector.

Figure 19

Experimental results of environmental adaptability. (a) Cost curve; (b) contact force; (c) Cartesian stiffness schedules; (d) Cartesian damping schedules.

Figure 20

Experimental comparison results. (a) Force control performance; (b) target stiffness; (c) target damping.

Figure 21

(a) The cost curve of the learning process. (b) Comparison of learning speed with other learning variable impedance control methods.

Figure 22

Computational time for each rollout.