Periodic Reporting for period 1 - REMARO (Reliable AI for Marine Robotics)
Periodo di rendicontazione: 2020-12-01 al 2022-11-30
In the first two years of operation, the ESRs of the network have produced methodological and conceptual results. Working with marine robotics and inspection tasks resulted in the creation of benchmark datasets for underwater computer vision and 3D scene reconstructions. Furthermore, we deployed an underwater drone in the canals of Copenhagen to gather real underwater data while imitating wall inspection missions. These new data sets will complement the closed data of industrial partners and will allow for open dissemination of results. The datasets can be used within and outside the network, to boost research on underwater perception.
Several test beds reflecting underwater inspection scenarios have been created as well, building upon the Robot Operating System (ROS), Gazebo simulator, and the uuv_simulator package for physics simulation. Visualization is being addressed using the engines of gazebo, airsim, Blender, and Unreal.
Besides perception, the REMARO ESRs continued to work on cognitive architectures for underwater robots using technologies like meta-control, ontologies, and planning.
While perception allows the underwater robots to "see", a cognitive architecture allows robots to make decisions, where to move and what actions to perform.
REMARO boasts three published papers, three accepted papers, and two submitted papers already. What is more important is that most research work exploits the diversity of the ESRs - multiple teams combine skills from robotics, artificial intelligence, and formal methods to achieve research breakthroughs. The network has organized a workshop at IROS 2021, has participated in the EU Robotics meeting, and maintains a healthy social media presence. Its training events are designed to involve industrial participation among the audience and as lecturers.
The expected results of the project involve several testing frameworks for perception components, verification methods for planning and failure recovery components, methods for increase of reliability through censor fusion, and methods for uncertainty assessment for machine learning components. As experience with artificial intelligence does not directly translate between domains, it is crucial that all these methods will be evaluated on submarine robotics applications and scenarios. This will enable a much faster adoption of AI in this sector, which in turn will allow new scientific studies of the oceans, and new safety and maintenance methods for underwater infrastructure.