Acting on Push Affordances: Adapting Dynamic MovementPrimitives Based on Object Behaviour

Senka Krivic, Emre Ugur and Justus Piater

Abstract— Nonprehensile manipulation such as pushingcan play a significant role in complex scenarios. Objectsmay have diverse, even anisotropic properties under push-ing in different environments. This increases the com-plexity of the pushing problem. We propose an approachto adapting dynamic movement primitives (DMPs) basedon the observed object-motion behaviour and experiencedforces. We also investigate an alternative optimal controlbased technique that enable dexterous and adaptive manip-ulation using pushability and manipulability of objects.

I. INTRODUCTION

Pushing is often used to simplify or to improvethe precision of complex manipulative skills. If youimagine you are trying to place an object, for exampleto put a book on a shelf, or to carry out any tabletopmanipulation, in most cases you will place the objectapproximately, and then use pushing movements tobring it to the desired position and orientation. Also,pushing is commonly used to move objects out of theway or to make grasping other objects easier. Enablingsimilar pushing skills for robots can serve multiple pur-poses: correcting the placement of an object, clearingpaths and shoving objects into free space, maneuveringlarge objects or objects that are hard to grasp, enhanc-ing pick-and-place operations, stc.

In our preliminary experiments, we observed thatsingle-contact pushing of objects even in a straightline is not trivial due to differing properties of objects.Humans are able to push various objects in a desiredway by pushing them approximately into designateddirections and changing the hand motion based onobserved senses and behaviours. We aim to incorporatea similar concept into the robot manipulation skills byadapting robot motion controllers based on the objectbehaviour and experienced forces. For motion controlwe use dynamic movement primitives (DMPs) whichwere proposed as an efficient way for learning andcontrolling complex robot behaviours [1].

*The research leading to these results has received fundingfrom the European Community’s Seventh Framework ProgrammeFP7/2007-2013 (Specific Programme Cooperation, Theme 3, Informa-tion and Communication Technologies) under grant agreement no.610532, Squirrel

All authors are with Institute of Computer Science,University of Innsbruck, 6020 Innsbruck, Austria,Contact information: senka.krivic@uibk.ac.at,emre.ugur@uibk.ac.at, justus.piater@uibk.ac.at

Fig. 1. Object data set used for pushing tests (left); pushingexperiments with a KUKA LWR arm (right)

II. RELATED WORKMany researchers have studied nonprehensile ma-

nipulation by pushing with robotic manipulators orwith the robot base. This topic raised interest in dif-ferent research areas which resulted in a variety of ap-proaches and problems related to pushing. Early workon pushing done by Mason and Lynch [2] presented atheoretical model of the dynamics of pushing with therobotic manipulator. One of the first systems that tookinto account vision feedback for building a forwardmodel of a planar object was created by Salganicoffet al. [3]. In later work, the action of pushing wasoften used for learning dynamic models, behavioursand object specific properties by observing results ofthe push [4], [5]. Several methods treat the problemof learning contact locations for successful pushes [6],[7]. Most of them focus on extracting the shape of theobject and determining contact locations that are goodfor pushing.

Existing algorithms have several limitations whichinclude the following. (i) Restriction to objects andenvironments which possess specific properties suchas quasi-statically sliding on a high-friction surface.(ii) Pushing planners are usually based on geometricproperties of objects and designed for specific sce-narios. (iii) Pushing to the goal position is done inclutter-free environments. (iv) Many methods require atime-expensive learning phase depending on the objectshape.

III. PROPOSED APPROACHObjects may have diverse properties under pushing

in complex scenarios. If you imagine a task, such ascleaning a child’s room, a robot has to be able tomaneuver diverse objects with different masses andfriction distributions. In tests of simple, straight-line

pushes which were done for a set of various objects(see Fig. 1), we observed that object loss is mostly de-pendent on the friction force and the object behaviourdynamics. Different object properties are manifestedunder different pushing dynamics and environmentconditions. For example, objects can flip due to highfriction or fast movements. In our pushing task, therobot is assumed to have no previous knowledge ofthe object. By pushing the object, it is possible to buildan object behaviour model on the fly based on obser-vations of object-environment dynamics. This model isthen used to adjust the robot motion controller. Theproposed framework is shown in Fig. 2..

A. Adjusting DMPs

Choosing suitable control parameter values for push-ing is frequently object-dependent. Instead of tuningcontrol parameters with respect to the object class orthe shape, we propose identifying the object behaviourby sensing and introducing corrective actions in a push-ing task. At the onset of a manipulation, the object isassumed to have ideal properties for pushing, movingonly together with the robot. During the executionthese assumptions are updated based on the objectdisplacement or end-effector force values with respectto the dynamics of the robot movements. Control pa-rameters are adapted based on the pushing experienceto provide desired object movements. Online modu-lation of dynamic movement primitives is done byadding coupling terms. These terms alter a primitivebased on intermediate success of object behavior andexperienced forces:

τ z = αz(βz(g − y)− z) + f(x) + IO (1)

τ y = zτ, x =−αx

1 + αeJe

where:

IO = Ktpp(E [xO]− xO) +KtF f(E [F |emin]− F )

Je =

∫ tc

t0

d (pd (γ) , xO) dt

Ktp = f(var(xO), E(xO)) and KtF = f(∇F, F ) are time-varying task sensitivities with respect to object pose xOand force F .

B. Optimal Control

As an alternative method, we propose an optimalcontrol strategy with respect to the pushing objectives.Objectives which define the cost function are as fol-lows: (i) The object should be delivered to the targetpose as fast as possible within given tracking-accuracyconstraints; (ii) Non-smooth and jerky movements ofthe robot should be minimized:

J = ct

∫ tf

t0

t dt+ cj

∫ tf

t0

N∑i=1

(d3xidt3

)2

dt (2)

adjustment mechanism

controller system environment

parameters

model behaviour

identification

model uncertainty

model identification

uncertaintyFig. 2. The proposed model behaviour identification adaptivecontrol approach for push-manipulation

Here, tf is a free variable representing the final pointin time, and jerk is defined in Cartesian coordinates xi.The weights ct and cj are coupled coefficients depend-ing on the trade-off between the time limit and desiredsmoothness of the robot movement. The constraints forpushing are first-order robot and object-environmentdynamic constraints together with initial conditions.In cluttered environments an additional constraint isgiven by minγ d(E [PO(t+ δt)] , pd(γ)) < ρ where d(·) isthe perpendicular distance of an object to the desiredcollision-free path pd(γ), E[·] is the expected valueover object movements, and ρ is the tolerance for pathtracking.

IV. CONCLUSIONExploiting object manipulability in complex scenar-

ios enables new capabilities and increases adaptivenessof a robot. A robot has to recognize specific objectbehaviours entailing strategies and the adaptation ofpushing control. At the same time, movement primi-tives should be adapted using available sensor infor-mation. Such an adaptive control approach for push-manipulation is presented in this paper.

REFERENCES

[1] Stefan Schaal, Jan Peters, Jun Nakanishi, and Auke Ijspeert.Learning movement primitives. In International Symposium onRobotics Research (ISRR2003). Springer, 2004.

[2] Kevin M. Lynch and Matthew T. Mason. Stable pushing: Me-chanics, controllability, and planning. The International Journal ofRobotics Research, 15(6):533–556, 1996.

[3] Marcos Salganicoff, Giorgio Metta, Andrea Oddera, and GiulioSandini. A vision-based learning method for pushing manipula-tion. In In AAAI Fall Symposium Series: Machine Learning in Vision:What Why and, 1993.

[4] D. Katz and O. Brock. Manipulating articulated objects withinteractive perception. In Robotics and Automation, 2008. ICRA2008. IEEE International Conference on, pages 272–277, May 2008.

[5] F. Ruiz-Ugalde, G. Cheng, and M. Beetz. Fast adaptation foreffect-aware pushing. In Humanoid Robots (Humanoids), 2011 11thIEEE-RAS International Conference on, pages 614–621, Oct 2011.

[6] T. Hermans, J.M. Rehg, and A.F. Bobick. Decoupling behavior,perception, and control for autonomous learning of affordances.In Robotics and Automation (ICRA), 2013 IEEE International Confer-ence on, pages 4989–4996, May 2013.

[7] M. Kopicki, S. Zurek, R. Stolkin, T. Morwald, and J. Wyatt.Learning to predict how rigid objects behave under simplemanipulation. In Robotics and Automation (ICRA), 2011 IEEEInternational Conference on, pages 5722–5729, May 2011.