Project type: Explore Project

Re-Use of Robotic-data in Production through search, simulation and learning

Project Description

We investigate – in close co-operation with four Danish companies, the cluster organization Odense Robotics and the research platform MADE – underlying problems of data re-use on assembly in production. In contrast to other fields in AI, the potential of exploiting large data collections is not realized in robotics yet. We aim at analyzing the underlying scientific and technical challenges as well as associated legal and privacy issues by means of three half days meetings of university partners and companies, one public workshop, and the preparation of four deliverables. The potential of re-use of data of past robot executions in industrial robotics is not realized yet due to scientific, technical and IPR issues. In this project, we conduct – in close co- operation with key industrial partners – a first investigation into the underlying problems to prepare a larger project application which is grounded in today’s challenges faced by companies. The investigation is done by means of three meetings of university partners and companies, one public workshop – which is planned to be open to DIREC, Odense Robotics and MADE (Manufacturing Academy of Denmark, www.made.dk) participants, and the preparation of three deliverables which address the underlying scientific, technical, privacy and legal issues. The ReRoPro project operates in the crossfield of classical AI and industrial robotics and also aims at bridging between DIREC and MADE by exploiting synergies between these two large national consortia in the area of AI and Robotics. A fourth deliverable will be a grant proposal with significant company co-financing. ReRoPro gets support by Odense Robotics and MADE. In contrast to, e.g., computer vision or speech recognition, there are only very few examples where data has been exploited across different robotics applications. In the area of computer vision – in particular, in the field of deep learning in the last decade – efficient software-modules for, e.g., object recognition and human pose estimation have been developed. The same holds for speech recognition and other areas. The availability and exploitation of large amounts of data has been crucial for these successes. However, similar successes have not yet emerged in robotics. The main reason for that are the particularities of robot data and the close connection to hardware infrastructure, which make it more difficult to transfer experience from one setup to another. In addition, IPR issues are a potential hurdle since the data about objects and processes is of great value for the companies. Therefore, companies are hesitant to share such data with potential competitors. As a consequence of the lack of availability of such structured data, today’s robot applications have to be built up from scratch for every new task. That leads to long and expensive set-up times, and thus limits the use of robot solutions. This, in turn, pushes production to countries with lower salaries than Denmark. If, however, data of past executions could be made available in a reusable format, it could be harnessed in the future for the robot control methods, which could reduce the setup time of the application and subsequently makes it cheaper and broader. In this project, we perform initial steps to address the complex issues involved in the problems outlined above by (1) performing first steps in defining appropriate data structures for data storage and retrieval as well as a high-level architecture that allows for exploitation of robot data, (2) formulating a number of learning problems with our three industrial partners routed in the need of industry that will constitute the use cases addressed in future projects, and (3) investigate the IPR-issues connected to the sharing of sensitive data. By means of three half-day meetings of AI experts, robotic researchers and companies and a public workshop the problems at hand will be explored and the results will be put down in three deliverables addressing the three issues mentioned above. A larger project proposal will be a fourth deliverable.

Value Creation

Scientific value creation  Scientific insight into the problem of re-use of data in robotics: The project will gather experts in the field and establish knowledge on the problem at hand in a strong consortium that plans to work on the problem in the coming years. The main challenges faced will be analyzed and spelled out and first steps towards valid data structures will be made already during the project period and summarized in one public deliverable. Business value creation The formulation of a set of learning problems grounded in the needs of industry: There have been past attempts to address the problem of re-use of robot data in the context of the general problem of cognition (e.g., during the Cognitive System Calls a decade ago). An important difference of ReRoPro and these past projects will be that we will root the problem in tasks that are relevant for industry already now. These problems also have much lower complexity than the general cognition problem which increases the success chances. In one deliverable, we will define a small set of learning problems which, when successfully addressed by our generic approach, will generate large value for the companies involved. The learning problems will be also specified to a degree that we can estimate that our approach will succeed with high likelihood. Both is required for gathering larger funding for our approach in the future. The awareness of legal challenges connected to the re-use of data: A prerequisite for companies to become involved will be that their production data will be protected. There are different means to address these issues (keeping data local, abstracting data into models which are shared, etc.). The awareness of these issues and possible solutions thereof are crucial for assuring the required trust for the companies involved and is therefore a basis for the business value to be generated. Capacity building Shaping of a larger proposal on re-use of data in assembly with strong company involvement: The explorer project ReRoPro is supposed to be the basis for forming a consortium for a rather large national project proposal with significant company involvement which is supposed to be handed in already in 2022. National and international visibility of the DIREC consortium by means of a one-day workshop on the re-use of robot data: As part of the project, we will organize a DIREC-workshop with top-notch national and international experts on the re-use of robotic data. This will give the DIREC project national and international visibility. Exploiting synergies between the MADE and DIREC consortium: ReRoPro is positioned in the interface between AI and industrial robotics, The MADE-consortium (Manufacturing Academy of Denmark), the leading research platform in Denmark in industrial robotics, will support ReRoPro. By that, the two large national consortia DIREC and MADE, both funded by InnovationFonden Danmark, can exploit synergies between them. Capacity Building for Education  We will integrate the insights gained during the ReRoPro project in our robotic educations, both on the master and PhD level. Since we have already a number of master and PhD theses running that touch the topic addressed in ReRoPro, the project will facilitate the systemization of the data handling in these and future theses. Moreover, since data management and data re-use is becoming increasingly important for robot applications in industry, we will integrate this aspect more systematically in our robotic courses. ReRoPro will also inspire future master theses at SDU Robotics and Software and KU/AAU computer science.

Participants

Project Manager

Norbert Krüger

Professor

University of Southern Denmark
The Maersk Mc-Kinney Moller Institute

E: norbert@mmmi.sdu.dk

Aljaz Kramberger

Assistant Professor

University of Southern Denmark
The Maersk Mc-Kinney Moller Institute

Jakob Wilm

Associate Professor

University of Southern Denmark
The Maersk Mc-Kinney Moller Institute

Mikkel Baun Kjærgaard

Professor

University of Southern Denmark
The Maersk Mc-Kinney Moller Institute

Anders Lyhne Kristensen

Professor

University of Southern Denmark
The Maersk Mc-Kinney Moller Institute

Kenny Erleben

Associate Professor

University of Copenhagen
Department of Computer Science

Sune Darkner

Associate Professor

University of Copenhagen
Department of Computer Science

Thomas Dyhre Nielsen

Professor

Aalborg University
Department of Computer Science

Alvaro Torralba

Associate Professor

Aalborg University
Department of Computer Science