Research Papers

Feedforward Input Generation Based on Neural Network Prediction in Multi-Joint Robots1

[+] Author and Article Information
Jonathan Asensio

Department of Systems Engineering & Control,
Polytechnic University of Valencia,
Valencia 46022, Spain
e-mail: jonathan.asensio@gmail.com

Wenjie Chen

Department of Mechanical Engineering,
University of California,
Berkeley, CA 94720
e-mail: wjchen@berkeley.edu

Masayoshi Tomizuka

Department of Mechanical Engineering,
University of California,
Berkeley, CA 94720
e-mail: tomizuka@me.berkeley.edu

If the computing resource and the sensor configuration allow, the end-effector sensor can also be used online [11]. This paper, however, will address the conservative case where the end-effector sensor is for off-line training use only, which is preferred in industry due to the cost saving and the limited real-time computational power.

The dynamic model and the controller introduced later will be greatly simplified if the joint compliance is absent (i.e., KJ=,DJ=, and q=qm/N). The proposed controller may not be necessary in this special case since the problems (such as mismatched dynamics and the residual vibration behavior) arising from the joint compliance may not exist.

Quantification of the required computation resources depends on various factors, such as the actual robotic system, the implementation platform, and the tools that are used (e.g. compilers, optimization options, and variable data types).

Note that the LCS in Fig. 9 exhibits overcorrection (e.g., LCS error in static periods is about 180 deg phase difference from the Basic setting) because the static error has already been reduced by the correction of the dynamic error. This overcorrection will disappear as the LCS continues learning iteratively. The NNP does not have this behavior since it is turned off during the static periods. However, this switching (around 8.9sec) causes the Cartesian space performance change of NNP (making it worse than LCS).

1 This work was supported by FANUC Corporation, Japan, and the UPV PROMOE exchange program, Spain. This paper was presented in part as the paper (DSC2012-8726) [1] in the 2012 ASME Dynamic System Control Conference.

2Corresponding author.

Contributed by the Dynamic Systems Division of ASME for publication in the JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL. Manuscript received November 11, 2012; final manuscript received November 2, 2013; published online January 20, 2014. Assoc. Editor: Won-jong Kim.

J. Dyn. Sys., Meas., Control 136(3), 031002 (Jan 20, 2014) (9 pages) Paper No: DS-12-1367; doi: 10.1115/1.4025986 History: Received November 11, 2012; Revised November 02, 2013

Learning feedforward control based on the available dynamic/kinematic system model and sensor information is generally effective for reducing the tracking error for a learned trajectory. For new trajectories, however, the system cannot benefit from previous learning data and it has to go through the learning process again to regain its performance. In industrial applications, this means production line has to stop for learning, and the overall productivity of the process is compromised. To solve this problem, this paper proposes a feedforward input generation scheme based on neural network (NN) prediction. Learning/training is performed for the NNs for a set of trajectories in advance. Then the feedforward torque input for any trajectory in the predefined workspace can be calculated according to the predicted error from multiple NNs managed with expert logic. Experimental study on a 6-DOF industrial robot has shown the superior performance of the proposed NN based feedforward control scheme in the position tracking as well as the residual vibration reduction, without any further learning or end-effector sensors during operation.

Copyright © 2014 by ASME
Your Session has timed out. Please sign back in to continue.



Grahic Jump Location
Fig. 3

NN predictor structure

Grahic Jump Location
Fig. 1

Robot control structure with reference and torque update. The ultimate objective is to make robot plant output qℓ track the load side reference trajectory qℓr. Only motor side position output qm is available for the real-time feedback. rq and τnl are the additional reference and feedforward torque updates to further compensate for the joint flexibility and the fictitious disturbance d, respectively. See Sec. 2.2 for more controller details.

Grahic Jump Location
Fig. 2

Control diagram with NN predictor

Grahic Jump Location
Fig. 4

6-DOF robot example (a) axis direction convention and (b) robot home position

Grahic Jump Location
Fig. 5

FANUC M-16iB robot system

Grahic Jump Location
Fig. 6

Training (blue) and validation (red) trajectories for joint 3

Grahic Jump Location
Fig. 7

Joint 5 model following error. Color: Red = 4, Blue = −4, Green = 0 in [rad/s2]. (a) and (b) are 2D and 3D distributions, respectively, based on joint 5 reference before preprocessing stage. (c) and (d) are 2D and 3D distributions, respectively, based on joint 5 motion influence (the reference combination from all joints) after preprocessing stage.

Grahic Jump Location
Fig. 8

Cost function convergence and parameter adaption over iteration during NN training process

Grahic Jump Location
Fig. 9

Experimental performance comparisons for error reductions. (a)–(c) the model following error on joint 2, 3, and 5, respectively. (d) the end-effector position error (in Euclidean distance, eX2+eY2+eZ2).




Some tools below are only available to our subscribers or users with an online account.

Related Content

Customize your page view by dragging and repositioning the boxes below.

Related Journal Articles
Related eBook Content
Topic Collections

Sorry! You do not have access to this content. For assistance or to subscribe, please contact us:

  • TELEPHONE: 1-800-843-2763 (Toll-free in the USA)
  • EMAIL: asmedigitalcollection@asme.org
Sign In