Scientific challenges to be addressed by the ERA Chair holder


The ERA Chair holder will be given full autonomy and independence in the design and implementation of his/her research strategy, but he/she should address the scientific challenges needed by LAMOR:

  • AI based robot learning (including Machine Learning, Deep learning, Reinforcement Learning, Continuous Learning, Data Pre-processing, Symbolic AI for Learning, Autonomous Learning and Multi-agent Learning) plays a pivotal role in all three LAMOR research domains. Robotic perceptioncan be significantly boosted by means of Machine Learning. Beyond avant-garde tasks approached via Deep Learning models (e.g. segmentation, classification, ...), of special importance is continuous robotic learning, i.e. the development of robotic models capable of learning to perform perception tasks while on operation, discovering and harnessing latent interrelationships between the tasks should they evolve or change over time. This adaptability requires importing ideas from transfer learning, evolutionary multitasking and representation learning. Robotic actuation(motion planning 5and control, interaction) also, in which the goal is to make robots autonomously learn how to solve a task that requires motion and interaction with the environment. Specifically, Machine Learning methods receive a feedback signal from the environment (reward) that helps them progressively find the best motion and interaction policy for the given task.
  • Reasoning, optimization and planning are key elements for supplying robots with cognitive functions, which can in turn enhance their capability for autonomous operation in diverse and flexible conditions. The incorporation of intelligent decision agents, expert systems, evolutionary algorithms and programmable automata for the optimization of robotic systems follows an active trend in the industrial environment regarding robotics. Such components of AI have as their main function to control robot systems both independently and in cooperation with other agents, industrial machines, scheduling systems, e.g. for deliveries or maintenance, as well as teams of individual robot units that need to coordinate their activities to dynamically solve complex manipulation tasks. AI planning takes an initial state and creates an action plan that seeks to achieve an objective. For the implementation of this functionality, software planners are used. However, these planners do not themselves transmit instructions to the hardware –an intermediate layer is usually required between the planner and the robots that can interpret each step of the plan and translate this into robot code executing its corresponding primitive. Reasoning is the stage that takes care of this interpretation and provides the robot with intelligent behavior via a processing architecture that allows it to learn and reason about how to behave in the face of complex objectives, even in arduous and dynamic environments.

To foster synergy with the existing LAMOR expertise the following research topics of the ERA Chair and his/her team are of the outmost importance:

  • AI for robotics applications that understand the context and environmentin which they operate (the robot understands besides just knowing where it is), and over time build underlying explanatory models that allow them to characterize real world phenomena. The algorithms should correctly classify information into proper categories, or make correct predictions, recommendations, or decisions based on data or models, and when unable to do so, report a measure of the pertaining error uncertainty.
  • Exploiting the power of representation learningoffered by AI end-to-end approaches, while at the same time leveraging the efficiency, robustness and interpretability offered by model-based methods. The decades of model-based research should not be dismissed, especially when offering the chance to achieve an interpretable AI system operating at high level of performance.
  • Validation and evaluation methods for AI for robotics, since it is critical that the results of such an AI system are reproducible, as well as reliable. A responsible and trustworthy AIsystem is one that works properly with a range of inputs and in a range of situations. Reproducibility describes whether an AI experiment exhibits the same behavior when repeated under the same conditions.






This project has received funding from the European Union's Horizon 2020 research and innovation programme under the Grant Agreement No. 952275