Implementation of Torque-based Control onIndustrial Robot for Human-Robot Cooperation

Seho Shin*, Yisoo Lee*, and Jaeheung Park*,**†

*Department of Intelligent Convergence SystemsSeoul National University, Korea

***Advanced Institutes for Convergence Technology, KoreaEmail: {shinsh, howcan11, park73}@snu.ac.kr

Abstract—Human-robot cooperation is getting impor-tant in nowadays factory environments, where safety isone of the most important factors. The use of torque-based controllers is one approach to provide safety andperformance in these applications. In this paper, theimplementation of torque-based control on industrialrobots are presented. First, dynamic properties of therobots, such as inertia and friction parameters, are iden-tified to be used in torque-based controllers. The torque-based controllers are then implemented on an industrialrobot, HA020. The performance of torque-based control isdemonstrated by the experiments of gravity compensation,joint space control, and the operational space control.

Index Terms—torque-based control, industrial robot, systemidentification

I. INTRODUCTION

Human-robot cooperation is one of the significant appli-cations in factories for assisting human manipulating heavyloads and teaching the robot complicated tasks. The function-ality of human-robot cooperation is typically implementedby installing force sensors to detect the load and human’sintention on an existing position-controlled robot [1]. How-ever, this approach has limitation in safety and performancebecause it does not provide inherent compliance and does notaccount for dynamics of the robot [2]. In order to improvethe performance and cost of the system, torque-based controlhas been considered for human-robot cooperation task. Oneof the main advantages of using torque-based controllers isto improve compliance and utilize the dynamics of the robot[3].

In this paper, implementing torque-based controller on anindustrial robot is discussed. The HA020 robot (see Figure 1)is used as a testbed, where only motor specification and DHparameter information were available. Therefore, the robotwas disassembled to obtain the dynamic properties of eachlink of the robot such as mass and inertia of each link [4].Joint friction is also experimentally identified since compen-sation of joint friction greatly affects the performance andstability of the robot [5]. With the estimated dynamics proper-ties and joint friction of the robot, the performance of torque-based controllers is evaluated by the following experiments;gravity compensation, joint space control, and the operational

†Jaeheung Park is the corresponding author.

Fig. 1. Hyundai HA020: it is a 6 DOF industrial manipulator. Yellowcircular arrows represent the rotational direction of each joint, and the redcircle is the gripper. The gripper has one joint, which is controlled separatelyby an external controller.

space control. The experiment of gravity compensation is toverify the robot’s dynamic properties, and system’s abilityto command torque on each joint instead of commandingdesired joint angles. The other two experiments are to developa control scheme for human-robot cooperation. Feasibility oftorque-based control on human-robot cooperation system isdemonstrated by these experiments.

II. HARDWARE SPECIFICATION

The Hyundai HA020 is a 6-DOF industrial robot, and itstotal weight is 240kg with a maximum payload of 20kg. TheHA020 robot uses AC servo motors with 17bit incrementalabsolute encoder. The gripper weighting 26kg is attached tothe robot, and it assists human robot cooperation. The gripperis controlled via an external control module.(see Figure 1)

The motion control and the servo driver part are mod-ified using the Synqnet-PCI motion controller in order tocommand torques to individual joints. Instead of the robot’steaching pendant, the robot can be controlled from Win-dows based computer with real-time robot control software,called RoboticsLab [6]. Roboticslab has physics-based re-altime robot control module and programmable interfacethat compute the robot’s dynamics and control torques. This

Fig. 2. Modified System Architecture for HA020

system uses a RTX(real-time extension) runtime to guaranteedeterministic time on Windows platform [6]. Therefore, thereadings of the joint angles from the encoder are passed tothe RoboticsLab, and the control module in the RoboticsLabcomputes torques and then commands them to each joint ofthe robot in real-time.(see Figure 2)

III. SYSTEM IDENTIFICATION

A. Link Properties

System identification of the robot is conducted to obtainthe dynamic properties of the robot. In order to calculatedynamics of a robot, the properties of mass, inertia, size, andfriction are necessary [7]. (see Figure 3) Since these parame-ters are not available for HA020, the robot is disassembled toobtain the mass and the center of mass of each link. With themeasured values, the moment of link inertia is estimated bya 3D model. In addition, rotor inertias are obtained from thespecification of each motor and accounted for in the dynamicmodel [8].

Fig. 3. Dynamic parameters for system identification

B. friction compensation

The industrial robot has large joint frictions due to the useof large motors and high gear-ratio. Therefore, high gains aretypically applied to compensate for the large joint frictionsin controlling the robot motion. This may adversely affect onthe system’s flexibilities, time delays, and sampling rate. Inimplementing torque-based control, the Coulomb and viscousfriction models are used and its parameters are estimated tocompensate for the friction directly. The theoretical frictioncompensation torque τ f ric is

τ f ric = fcsign(q)+ fvq (1)

where fc and fv are Coulomb and viscous friction parameters,respectively.

In practice, the direction and magnitude of τ f ric maychange frequently and abruptly when the joint velocity, q,is near zero. Another problem with the friction compensationis when the robot starts to move. When the velocity of thejoint is zero, the direction of Coulomb friction cannot be de-termined. Due to this problem, the robot cannot start to moveuntil the position error accumulate to a large value. Once therobot begins to move, it overshoots the desired trajectory. Toovercome these problems, the friction compensation torqueis designed as the following when the joint velocity is belowa certain threshold, qt . In order to determine the directionof friction compensation torque in the region of |q|< qt , thedesired input torque, τ , is used. If the command torque, |τ| isgreater than a threshold, τt , Coulomb and viscous frictionsare applied. In the region of |τ| < τt region, the frictioncompensation torque is linearly interpolated based on thedesired command torque, τ . (see Figure 4)

τ f ric =

fcsign(q)+ fvq if |q| ≥ qt

fcsign(τ)+ fvq if |τ| ≥ τt

fc − fc(τt−|τ|)τt

+ fvq if |τ|< τt

if |q|< qt

(2)

IV. CONTROL FRAMEWORK

To evaluate the performance of the torque-based con-trollers, the joint space control and the operational spacecontrol are implemented.

A. Joint space control

The dynamics of the robot in joint space is described by[9]

A(q)q+b(q, q)+g(q) = Γ, (3)

where A(q), b(q, q), and g(q) are the inertia matrix, Corio-lis/centrifugal torques, and gravity torques, respectively. Dy-namic decoupling is achieved by using the control structure

Γ = A(q)Γ∗+ b(q, q)+ g(q), (4)

where, A(q), b(q, q), and g(q) present the estimates of A(q),b(q, q), and g(q). The term, Γ∗, is the input for the decoupledcontrol, which is composed with PD control. In addition, thefeed-forward compensation torque is added for friction.

B. Operational space control

The dynamic behavior of a manipulator in the operationalspace is specified as follows [9]

Λ(x)x+µ(x, x)+ p(x) = F, (5)

where Λ(x), µ(x, x), and p(x) are the inertia matrix, theCoriolis/centrifugal forces, and gravity forces in the opera-tional space, respectively. The control force, F, in (6), canbe composed to provide a decoupled control structure bychoosing

F = Λ(x) f ∗+ µ(x, x)+ p(x), (6)

Fig. 5. Snapshot of the gravity compensation experiment. The robot is compliant to external forces during the gravity compensation. The user is able tochange the configuration of the robot by applying force on any link of the robot.

(a)

(b)

(c)

Fig. 4. (a) Friction compensation torques in case of interpolation and non-interpolation in the region of small velocity (b) Plot of input torque, τ , overtime. The dotted line represents the threshold of torque, τt (c) Plot of velocity,q, over time. The dotted line is the threshold of velocity, qt . The joint velocitystarts to increase at 3 seconds.

where, Λ(x), µ(x, x, and p represent the estimates of Λ(x),µ(x, x), and p(x). The term, f ∗, is the input for the decoupledcontrol. In order to achieve the desired motion in the oper-ational space, a linear dynamic behavior can be composedwith a PD controller of the form

f ∗ = kp(x− xd)− kv(x− xd), (7)

where, xd and xd are the desired position and velocity ofthe end-effector, respectively. kp and kv are the position andvelocity gains. Once f ∗ is determined, Γ can be produced by

Γ = JT F, (8)

where, Γ and JT are the command torque and the transpose ofJacobian matrix, respectively. The command torque is appliedto the robot with the feed-forward compensation torque forfriction.

V. EXPERIMENTAL VALIDATION

Experiments have been conducted to demonstrate the fric-tion compensation, gravity compensation, joint space controland the operational space control. The servo rate of thecontroller of the robot is set to 500Hz. For trajectory tracking,incremental displacements are specified using cubic splinealgorithm at each control cycle. For the velocity feedback,measured angles from encoders are fed to a low-pass filterwith a cutoff frequency of 10Hz.

A. Friction Compensation

Figure 6 compares the results between command torqueand friction-compensated command torque for the first jointaxis when the robot starts to move using the joint spacecontrol. According to Equation (1), the compensated torqueis interpolated in |q| < qt region. The dotted line shows theboundary of |q|= qt . With this compensation for friction, therobot was able to start smoothly without too much overshoot,which can be seen in the experiment of joint space control.

Fig. 6. Experimental result of the friction compensation. The frictioncompensation torque is applied by the friction compensation method.

B. Gravity Compensation

The gravity compensation is to verify whether the systemis able to command torque to each joint instead of command-ing desired joint angles. Also, it is to validate if the robot canshow compliant motion using the torque-based control.

Before applying gravity compensation torque, the robotcould not support its own weight. Figure 5 shows the gravitycompensation experiment. The robot is compliant to the ap-plied forces by the operator. From the experiment, it is verifiedthat the torque-based controller effectively compensates forthe gravity using the estimated dynamic parameters of therobot and the robot is able to perform the compliant motion.

C. Joint Space Control

All links are set to run a sinusoidal trajectory at a periodof 8 seconds and at an amplitude of 15 degrees. Figure 7shows the responses of the 4th joint axis. It shows the resultof tracking performance of the joint. The maximum absoluteerror is 0.0243 radian during the experiment.

13 14 15 16 17 18 19 20 21 22 23 24

1.3

1.4

1.5

1.6

1.7

1.8

1.9

Time(t)

Ang

le(r

ad)

Desired TrajectoryActural Trajectory

Fig. 7. Plot of measured joint angle and desired joint angle in the experimentof joint-space motion

D. Operational Space Control

The experimental results of the operational space controlis plotted in Figure 8. The robot is commanded to move0.10 meters along the x-axis. The maximum absolute erroris 0.0044 meters during the experiment.

E. Discussion

The above experiments demonstrate the feasibility ofthe torque-based control on an industrial robot for human-robot cooperation. The gravity compensation, the joint spacecontrol, and the operational space control are successfullyimplemented. However, the tracking performance of the jointspace and the operational space control shows the rooms to beimproved further. The friction does not seem to be well com-pensated in this study due to the insufficient modeling and itsparameter estimations. It is believed that improvement on thefriction compensation will greatly improve the performanceof these torque-based controllers on industrial robots.

11 12 13 14 15 16 17 18−0.04

−0.02

0

0.02

0.04

0.06

0.08

Time(t)

Pos

ition

(m)

Desired TrajectoryActural Trajectory

Fig. 8. The experimental result of the operational-space control. The desiredcommand(0.10 meters) is given to the robot along the x-axis.

VI. CONCLUSION

This paper presents an implementation of torque-basedcontroller for the industrial robot, HA020. Dynamic parame-ters have been identified by disassembling the robot and thefriction parameters are estimated through experiments. Theseparameters were experimentally validated through the gravitycompensation controller. Once dynamic parameters are veri-fied, the gravity compensation, the joint-space control, and theoperational space control are implemented. The experimentalresults demonstrate the capabilities of torque-based control onindustrial robots. This torque-based control is expected to beuseful in human-robot cooperation in providing complianceand high performance.

ACKNOWLEDGMENT

This work was supported by Seoul R&BD Program(No.JP100106).

REFERENCES

[1] K. Kosuge and N. Kazamura, Control of a robot handling an object incooperation with a human. IEEE International Workshop on Robotand Human Communication., 142-147, 1997.

[2] P. K. Khosla and T. Kanade, Experimental evaluation of nonlinear feed-back and feedforward control schemes for manipulators. InternationalJournal of Robotics Research., 7(1), 18-28, 1988.

[3] J. Swevers, W. Verdonck and J. De. Schutter, Dynamic model identifi-cation for industrial robots. IEEE Control Systems Magazine., 27(5),58-71, 2007.

[4] R. Jamisola, M. Ang, T. M. Lim, O. Khatib and S. Y. Lim, Dynamicsindentification and control of an industrial robot. IEEE InternationalConference on Advanced Robotics., 268-273, 1999.

[5] F. J. T. Vargas, E. R. De Fieri and E.B. Castelan, Identification andfriction compensation for an industrial robot using two degrees offreedom controller. International Conference on Control, Automation,Robotics and Vision., 1146-1151, 2004.

[6] SimLab, ROBOTICSLAB, Robotics SW Development Environment In-cluding Dynamics/Control Engines. http://www.rlab.co.kr, 2010.

[7] J. Swevers, W. Verdonck and J. De. Schutter, Motion Control ofIndustrial Robots in Operational Space: Analysis and Experiments withthe PA10 Arm. Advances in Robot Manipulators., 417-442, 2010.

[8] K. Radkhah, D. Kulic, and E. Croft, Dynamic Parameter Identificationfor the CRS A460 Robot. IEEE International Conference on IntelligentRobots and Systems., 3842-3847, 2007.

[9] O. Khatib, The Operational Space Framework. JSME InternationalJournal. Ser.C:Dynamics, Control, Robotics, Design and Manufactur-ing., 36(3), 227-287, 1993.