Learning Robust Dual-Arm Coordination in the Manipulation of Elastic
Elements from Intuitive Human Demonstrations
Cristian Vergara1 , Juan Sebastian Cuellar1 , Federico Ulloa1 , Santiago Iregui1 , Lander Vanroye1 , Ruan Viljoen1 Yudha Pane1 , Joris De Schutter1 , Erwin Aertbeliën1 , and Wilm Decré1
Abstract— This paper presents a framework that facilitates
and expedites assembly applications that involve dual-arm manipulation of flexible and fixtureless elements. This concept enables users without extensive expertise in robotics to teach the required motion, forces, torques, and arms coordination to perform these assembly operations, using devices that do not involve the robot platform. Subsequently, a motion/wrench model is learned using trajectory parameterized Probabilistic Fig. 1: User demonstrations for an o-ring installation (Upper). Our system Principal Component Analysis (traPPCA). Finally, the dual-arm generalizes the acquired signals, coordinating both arms motions to perform task is robustly performed by using the constraint-based con- the task (Lower). From left to right, the task steps are: initial configuration, troller expression-based Task Specification Language (eTaSL). o-ring picking, installation in the mock-up piston, and o-ring release. The installation of o-rings in pistons is presented as an example of an industrially relevant use case. 4) Reactive Controller: The dual-arm motion is deliv- I. I NTRODUCTION ered by combining learned and modeled information in our One of the main challenges in agile manufacturing is the constraint-based reactive controller eTaSL [5]. In this way, capability to facilitate and expedite the deployment of robot robustness against uncertainties associated with imprecise applications that involve the manipulation of elastic elements sensors and system calibration is increased. [1], e.g., the installation of o-rings in rigid objects [2]. This III. R ESULTS application brings challenges associated with the required A motion/wrench model is trained from five demonstra- motions, forces, and coordination of dual-arm manipulators tions, encoding variability associated with different relative to pick a elastic o-ring from a support, interact with a housing positions between the o-ring support and the mock-up piston. to install this o-ring in a specific grove, and perform the Fig. 1 shows four corresponding instances between one release motions (see fig. 1). The main contribution of this demonstration and an autonomous execution of the robot. paper is an extension of our work in [3] and [4] to learn and generalize dual-arm motion, force, torque, and coordination IV. D ISCUSSIONS AND C ONCLUSIONS to manipulate elastic elements. The proposed synergy between modeled and learned infor- II. A PPROACH mation enables a robust motion/wrench generalization while Our approach is described by the following technologies: dealing with changes in the relative positions between the 1) Application specific components: Additive manufactur- o-ring support and the mock-up piston. We believe this ing is used to develop application specific components (see approach has the potential to unveil a new era of automation, fig.1), easing the task execution for both the operator and the offering the possibility to automate industrial processes that manipulator, while increasing the application reliability. were not (technically and/or economically) feasible before, 2) Human interface: As in [3], task teaching devices while also dealing with small production series that have to equipped with pose and force/torque sensors enable intuitive be deployed quickly. task demonstrations. In this way, the user does not need to R EFERENCES deal with the robot body inertia, as in kinesthetic teaching, [1] K. Chatzilygeroudis, B. Fichera, I. Lauzana, F. Bu, K. Yao, F. Khadivar, and has real contact feedback. and A. Billard, “Benchmark for bimanual robotic manipulation of semi- 3) Learning Methodology: TraPPCA [3] is used to learn a deformable objects,” IEEE Robot. and Autom. Lett., 2020. motion/wrench model from the data acquired during demon- [2] I. G. Ramirez-Alpizar, K. Harada, and E. Yoshida, “Human-based framework for the assembly of elastic objects by a dual-arm robot,” strations. In this way, our system can robustly generalize the ROBOMECH Journal, 2017. approach and contact information corresponding to dual-arm [3] C. Vergara, E. Aertbeliën, and J. De Schutter, “Learning robust manipulation of elastic elements from user demonstrations. manipulation tasks involving contact using trajectory parameterized probabilistic principal component analysis,” in Proc. IEEE/RSJ Int. 1 Robotics Research Team, RAM Division, KU Leuven, Bel- Conf. Int. Robot. and Syst., 2020. gium, Core Lab Flanders Make, http://www.mech.kuleuven. [4] C. Vergara, J. De Schutter, and E. Aertbeliën, “Combining imitation be/robotics. This work belongs to Chorrobot, winner of the learning with constraint-based task specification and control,” IEEE ”KUKA Innovation Award 2021 - AI Challenge”, which was sup- Robot. and Autom. Lett., April 2019. ported by ERC Advanced Grant: ROBOTGENSKILL and Flanders Make [5] E. Aertbeliën and J. De Schutter, “ETaSL/eTC: A constraint-based task Projects: PROUD, PROROB and MULTIROB. Corresponding author: specification language and robot controller using expression graphs,” in cristian.vergara@kuleuven.be Proc. IEEE/RSJ Int. Conf. Int. Robot. and Syst., 2014.