In the last two months our project activities were three-folded:

• We wrote two papers. One targeting at the Human-Robot Interaction community where we presented results from our user study with Harting worker. The other paper presents our newly developed controller, the Assisted Gravitation Compensation Mode, to the robotics community. Further details are displayed bellow.
• We have established our requirements and specifications for the perceptual components and have started to integrate some existing work.
• We have given quite a number of FlexIRob demos disseminating the FlexIRob ideas.

Dissemination Activities

 In the last two months we demonstrated our FlexIRobs system several times mostly in the context of educational workshops. Two companies which held their internal workshops in the rooms of the university took the opportunity to get an impression of future methods in human-robot co-worker scenarios. Further, FlexIRob was part of the professional orientation course for students carried out by the teutolab.

Obstacle Avoidance Requirements and Specifications

In our FlexIRob system we will focus on detecting moving/movable obstacles. We will rely on the Articulated Scene Modeling approach which we have developed in previous projects. We selected the Articulated Scene Model because it allows to distinguish model-free between static background and dynamic foreground and because it fits the FlexIRob paradigm best.

Submitted publications

"Assisted Gravity Compensation to Cope with the Complexity of Kinesthetic Teaching on Redundant Robots"

• Current State: submitted to the 2013 IEEE International Conference on Robotics and Automation (ICRA 2013) on 17th September
• Content:
  The scope of this paper is to propose a new interaction control concept for teaching redundant robots with kinesthetic teaching. We thereby focus on the technical aspect and explain our new Assisted Gravity Compensation mode in detail.

Efficient reconfiguration of advanced robot systems to new tasks or environments is an ongoing research challenge. While imitation learning methods as well as modeling and simulation tools are continuously improving to reduce the inherent complexity of reconfiguration, novel robot systems are entering the field which pose new challenges. The paper addresses one important challenge in this area, which is the programming of kinematically redundant robots.

An important result of our field study with 48 industrial workers at Harting is that standard programming-by-demonstration methods for teaching task-space trajectories on a redundant robot using physical human-robot interaction are too complex for non-expert human tutors because of the complexity of simultaneously considering the redundancy resolution and the actual task specification. We report that most of the participants utilizing standard programming-by-demonstration methods were not able to accurately and collision-free follow a styrofoam parcours in a confined workspace.

We therefore propose a new interaction concept for redundant robot systems, Assisted Gravity Compensation, based on a hierarchical control scheme, separating task-space programming from the redundancy resolution. We propose to split the entire task into the two aforementioned aspects of kinesthetic teaching, namely task-space control of the end-effector (world mode) and joint-space control (redundancy resolution) and to assist the user in the resolution of the redundancy for letting him focus on the actual kinesthetic teaching task. The idea is to allow free movement in the taskspace by tracking the movement given by the user through kinesthetic teaching, while simultaneously controlling the joint-space according to the constrained environment. For this interaction concept we identify three requirements: a compliant and redundant
robot platform enabling close and intuitive human-robot interaction, a redundancy resolution module providing the user with a valid redundancy resolution of the inverse kinematics and a hierarchical controller fusing the cartesian task from kinesthetic teaching in taskspace and the joint configuration given by the redundancy resolution.

This interaction concept is implemented by extending our development platform FlexIRob utilizing robot platform KUKA Lightweight Robot IV. The redundancy resolution is provided by a previously trained Neural Network component (taught kinesthetically) with an encoded inverse kinematics to comply to the constraints imposed by the environment to avoid collisions with the static scene. In addition, we use a hierarchical position controller, based on ideas by Grupen and Huber (CBF controller).

A short evaluation of the new Assisted Gravity Compensation control mode by means of the results of the field study shows that the complexity of a kinesthetic teaching task is reduced, which is revealed by an improved task performance and significant less collisions. In addition, analysis of the questionaire reveals an improved user experience. In particular, a highly significant effect on the simplicity of handling the robot during the teaching task is observable.

The new Assisted Gravity Compensation mode allows non-expert users to perform kinesthetic teaching tasks faster and with higher precision making kinesthetic teaching an efficient programming-by-demonstration method for redundant robots.

"User Study on Kinesthetic Teaching of Redundant Robots in Task and Configuration Space"

• Current State: initially submitted to the Journal of Human-Robot Interaction - Special Issue: HRI System Studies on 31th August; submitted revised version on 29th October
• Content:
  The focus of this paper is the Harting study on pHRI. We give a comprehensive motivation and description of the study on kinesthetic teaching of redundant robots and present a profound report of the observations, results and insights we gained.

The recent advent of compliant and kinematically redundant robots poses new research challenges for human-robot interaction. While these robots provide high flexibility for the realization of complex applications, the gained flexibility generates the need for additional modeling steps and the definition of criteria for redundancy resolution constraining the robot’s movement generation. The explicit modeling of such criteria usually require experts to adapt the robot’s movement generation subsystem. A typical way of dealing with this configuration challenge is to utilize kinesthetic teaching and guide the robot to implicitly model the specific constraints in task and configuration space.

Current approaches to kinesthetic teaching do not reflect the standard technical approaches to redundancy control, which typically separate task space planning and constraint resolution in configuration space. Furthermore, dedicated interaction support for kinesthetic teaching of constraints in confined spaces has not yet been considered. That is, standard robot control as we also apply it in this contribution distinguishes task and configuration space in the redundancy resolution, but users have to deal with the complexity of simultaneously considering the redundancy resolution in configuration space and the specification of trajectories in task space.

Based on these considerations, we propose to comply interactive teaching process to this logical distinction between configuration and task space, which allows to decompose the overall adaptation process into two separate but consecutive phases: The first phase is the Configuration Phase where kinesthetic teaching is used to implicitly model obstacles or task-independent constraints which are relevant for movement generation in the joint space of the robot. Here, constraint knowledge is implicitly transferred through application of a learning method from the human tutor to the robot system without explicit constraint modeling. The learned inverse kinematics mapping is embedded in a hierarchical controller, allowing for execution of arbitrary motions in task space but respecting the learned constraints.

The second phase is the Programming Phase where the human tutor kinesthetically teaches a specific task-space trajectory just by applying forces to the end-effector.
The robot’s joints are in this phase controlled by the hierarchical controller trained in the previous phase. This allows the users to program a specific task without consideration of the environment. We refer to this control mode as Assisted Gravity Compensation.

As a primary goal of this contribution, we evaluated in the Harting study whether kinesthetic teaching of constraints in configuration space is possible and if a separation of task and configuration space interaction provides benefits compared to the typical teach-in methods which do not consider this separation. A secondary goal of the conducted user study has been to assess if this system concept, namely explicitly distinguishing between configuration and task space interaction, allows to reduce the complexity in adapting a redundant robot to new  environments and tasks.

In order to briefly summarize the results of the analysis of the study here, we report that

• concerning the general experience with our FlexIRob system, the physical interaction is easy and comfortable, not threatening to the users and self-explanatory in particular for people with good spatial vision abilities.
• concerning the configuration phase, our expectations in terms of

1. time required for reconfiguration,
2. collision avoidance in joint space and the
3. task-space accuracy

of the learned hierarchical controller were not entirely confirmed. Nevertheless, the objective results show that the configuration of the redundancy resolution through kinesthetic teaching and the learning of the inverse kinematics mapping can be done in less than 2 minutes and that at least half of the users could perform the kinesthetic teaching accurately such that collisions with the environment are avoided.