IKTPLUSS-IKT og digital innovasjon

The future of industrial robotics is rooted in the development of robots that can collaborate and learn with humans. These collaborative robots would have the ability to evolve and improve their behaviors through the usage of machine learning algorithms. However, understanding how to control and test the learning skills of uncaged, single- or multi-arm robots and their ability to safely interact with humans is challenging as their expected improvements is not precisely known. Testing such robots is becoming a crucial research area where the combination of expertise in software testing, machine learning and robotics is strongly required. The ambition of the multi-disciplinary T-LARGO project is to develop a new scientific and technological foundation enabling the testing of learning collaborative robots. Its main objective is the construction of an open test platform dedicated to collaborative robots while its impact lies in major scientific breakthroughs on how to test and control robots equipped with artificial intelligence. In 2019, we explored in depth the challenges raised by the testing of machine learning and deep learning models and started to design a methodology named DeepRegression, that we experiment on a UR3 collaborative robotic platform. In 2020, we designed a methodology named MeTeRo in which we explore different metamorphic relations that are created to test the robotic planners and how the UR3 collaborative robot behaves when deployed in an environment where humans are interacting with these cobots. In another line of work, we used generative adversarial networks (GAN's) to generate new samples in a live environment to minimize scenery misclassifications by cobots in a continuous integration environment. Following our results obtained on constructive disjunction, we also investigated how constraint acquisition can be used to learn disjunctive schedules, typically used in robot tasks planning and scheduling. In 2021, in addition to publishing and disseminating the research results of the project, we conducted research on using constraint programming for itemset mining (which can be useful for test practices); we developed a new constraint acquisition algorithm, named GECQA, for learning qualitative constraint networks in the context of scheduling robotic tasks; and we proposed a new algorithm that combines constraint acquisition with optimization. We also created a novel dataset using a RGBD camera and the robotic arm UR3. This dataset is useful to detect the safe and unsafe environments for humans to work with robots and to create relations between images and an ontology based on the environment. The relations and the ontology both create a scene graph which will train GNN's to understand the robotic environment and act according to different situations. In another line of work, we explored the usage of GANs for class-wise data augmentation and spinal networks along with sharpness aware minimization techniques for training. In 2022, we finalized the research activities of the project by publishing the main results in high-impact conferences and journals. We also took advantage of the visit of our international partner (i.e., LIRMM France) to strengthen the collaboration and discuss of further activities. The project results have been instrumental to acquire new EU-funded projects (i.e., AI4CCAM, MARS) in the context of using AI for testing autonomous systems.

The T-Largo project scientific impact includes the publication of one book chapter in the prestigious Springer Nature collection, 2 journal articles and 10 international conference or workshop papers including one AAAI paper and one IJCAI article. Through a dedicated collaboration activity with ABB Robotics and the French Laboratory of Informatics and Robotics of Montpellier (LIRMM), we developed five software prototype tools that can generate test trajectories to stress test industrial robots using advance constraint reasoning, that can perform constraint acquisition to learn qualitative constraint networks, that implement robust image classification. These results contribute to the demonstration that symbolic machine learning technique such as constraint reasoning and acquisition are relevant for testing collaborative robots. The T-Largo project results have been instrumental to acquire two new EU-funded projects, namely AI4CCAM (TrustworthyAI for Connected Cooperative Automated Mobility) and MARS (AI-augmented digital manufacturing) that will both start early 2023. Acquiring these new HEU projects is impactful for Simula and will contribute to the European strategy of the research laboratory.