Study results

In the last two months our main focus was on analyzing the data that we collected with our FlexIRob system during the user study on physical human robot interaction at Harting. We were able to evaluate the questionnaire and got first insights about the perception of our system by potential co-workers.

Descriptive evaluation

We had 45 users, 16 male and 29 female, belonging to the working units 'cage assembly', 'mechanical processing', 'die casting', 'internal part production', 'machine assembly', and 'prefabrication'. Most of them were working for 1 to 33 years at Harting (M = 11.36 years, SD = 9.78 years, Min = 1, Max = 33). The participants' mother tongue distributes over German (77,8%), Russian (11,1%), Turkish (6,7%) and Greek(2,2 %). Concerning their highest level of education, almost a half of the participants declared the receipt of a certificate of secondary education (CSE) or a general CSE (Haupt- oder Realschulabschluss), ca. 13% received a vocational or university-entrance diploma, and the remaining participants finished a vocational education (Berufsausbildung).

Answering the questions raised

As reported earlier, we conducted the study to answer questions concerning the perception and manageability of the FlexIRob system. In particular: Is such a physical MMI actually intuitive? For all users? Is it comfortable? How exhausting is the work with the robot arm?  Is the reconguration of the robot practicable? How well and how fast can it be done?  What are the characteristics of the training data provided by the naive users of the system?

The results are very convincing. First of all, the users felt not threatened by the robot and perceived it to be reliable, which is an important fact regarding the applicability in future industrial co-worker scenarios. Concerning the physical interaction, they stated that the robot was easy and comfortable to handle as well as the operation of the robot to be self-explanatory. Additionally, they felt the system to give helpful feedback during the collaboration. 

Evaluation concerning assisted vs. not-assisted wire-loop game

As reported earlier, we divided the participants into two groups during the wire-loop game: One group (group A) was assisted by the robot that respects the constraints of the environment and the other one (group N) not. Each joint of the robot arm needs to be controlled manually. Our motivation was to evaluate the new control mode, i.e. whether solving the wire-loop game was easier for participants of group A or not. Again, the results are promising. First of all, we can report that there is a significant difference in the perceived simplicity of operating the robot arm. The participants of group A appraised the operation easier than the ones from group N. Second, group A declared the settings of the robot to be significant more appropriate than group N. Finally, we could detect a marginal significant effect on the handling of the robot arm during the interaction. Participants of group A felt the handling more comfortable than the ones from group N.

A first analysis of the trajectories taught-in during the wire-loop game shows differences in the duration and the smoothness of the trajectories between the two groups. The assisted group is significant faster in teaching a trajectory than the not-assisted group. Group A also produces much smoother and more comparable trajectories than group N.

The general idea of the MoFTaG experiment is to use physical interaction with a human tutor to teach important areas in the working places and corresponding specific configurations to the robot. Our hypothesis is that such man-machine interaction (MMI) provides an intuitive way to teach an inverse kinematic to redundant robots by users of such assisting systems without needing an expert for programming the required robot kinematics. Of course, such a hypothesis raises a lot of questions like: Is such a physical MMI actually intuitive? For all users? Is it comfortable? How exhausting is the work with the robot arm? Is the reconfiguration of the robot practicable? How well and how fast can it be done? What are the characteristics of the training data provided by the naive users of the system?

To answer these questions we conducted a study with workers of Harting (HARTING Subsidiary - Development and Production for Industrial Connectors: HARTING Electric GmbH & Co. KG). The study took place at Harting from 10th to 16th of April. The experiment for each participant was divided into three main parts: a warm-up phase, reconfiguration phase (3 trials), and the wire-loop game. During the warm-up phase we showed a video, that informs about the two different operational modes. Furthermore, the participants had 2 minutes to become familiar with the robot arm and the physical interaction. During the reconfiguration phase the participant was asked to train the FlexIRob system according to a constrained working space. Once the system was trained, a reference movement (a straight line) was performed by the robot arm to give feedback to the participants. This procedure was repeated two more times. Finally, the participant was asked to play a variation of the wire-loop game. The task was to move the robot with the gripper along a given show-jumping course in the constrained working space. This game should simulate a tech-in of a concrete task to the robot. We had two conditions: The first group had to perform the task without assistance by the system. Hence, they had to take care of all the joints in order to not collide with the obstacles. The second group just needs to move the gripper. The joint configuration is controlled by our system through a network trained specifically to the current environment. We had 49 participants - 31 female and 18 male. They belong to the units 'cage assembly', 'mechanical processing', 'die casting', 'internal part production', 'machine assembly', and 'prefabrication'. They are truly naive with respect to physical interaction with such robots. At most, they might know big industrial (6DOF) robots behind "glass".

Some first impressions on the study results are: Our software system and the Kuka arm were running very stable. There were no breakdowns and overheating even though our FlexIRob system was running 13 hours per day and 4 days in total. All participants were able to finish successfully the study task. We observed a wide variety of interactions with the Kuka arm. Everyone played the assisted wire-loop game without any troubles independent from the problems he/she had in the warm-up and reconfiguration phase. Finally, we had the impression that everyone liked to interact with the Kuka arm. In particular, they seemed to have a lot of fun when playing with the Kuka arm, bending all joints, and transforming it to extremely unusual postures.

Industrial co-worker scenarios require save, flexible, and efficient control of robots. The goal of MoFTaG is to transform a data-driven model-free learning method that has been recently developed for high DOF-humanoid platforms into a ready-to apply real world tool for teaching co-worker robots flexible motion. Concretely spoken, our compliant 7-DOF KUKA Light Weight Robot 4 with force-torque sensors in each joint explores the task space with a human tutor through kinesthetic teaching. This procedure provides training data for the RNN controller which provides constraints for the redundancy resolution in the trained areas and generalize to the untrained workspace areas. Therefore, this approach allows a flexible and fast reconfiguration of the arm to new environments.

As a first step, we have presented at CogSys 2012 a poster showing some results of the accuracy analysis of our system. The evaluation is based on 3D trajectories in task space generated between random points by our hybrid controller which has been enhanced with the learned constraints (taught on the real arm through physical interaction). Further, we plan to analyze the intuitiveness of the training and the quality of the training data in a study with workers from a company who might be potential addressees of our system. Our dissemination partner OWL Maschinenbau established a contact to Harting (www.harting.de) which could provide us with workers who normally assemble products for small-series productions. As their working places are often quite non-ergonomic, these workers might benefit in future from a third robotic hand.

Two important aspects need to be considered in the study design: first, the safety of our arm and second, the motivation for the study. For safety, our robot was assessed by the company's Safety department and the Employers Mutual Insurance association. Due to our different safety levels built into our system and the low stiffness of the arm during execution, we got the admission for our study. For the motivation selling the study, we have considered in particular the fact that no fears with respect to job loss should be induced. Details on the story and the planned questionnaires could be looked up in the attachments.

In co-worker scenarios, flexibility by allowing for changes in kinematic configuration, e.g. through application of new tools or new degrees of freedom in re-configurable robots, is highly desirable. Current industrial practice requires costly and tedious reprogramming by experts. To facilitate and speed up this inefficient process, we propose to use a model-free learning method that enables a non-expert user to record a limited number of data-points in task-relevant areas of the workspace. So far, we have derived a benchmark scenario including reconfigurable physical obstacles and 3D-sensing capabilities and are planning a first extensive user study with an industrial partner.