The key ingredients of Artificial Intelligence
The team competences cover the traditional loop “Learn-Reason-Interact” that constitutes the base of cognitive systems. These components are the key ingredients of Artificial Intelligence:
Learning
Data-driven predictions and decisions
Learning explores the study and construction of algorithms that can learn from and make predictions on data; such algorithms overcome traditional programming by making data-driven predictions or decisions, through building a model from sample inputs. In this context, the team is working on:
- Machine Learning
- Data Analytics
- Novelty Detection
- Diagnostic
Space missions deliver plenty of data, and the trend is that they will deliver even more in the future. Due to this large amount of data, we need intelligent ways to automatically analyze data. Our group specializes in advanced data analytics, visual analytics and machine learning to cope with information overload.
We used advanced data analytics to support anomaly detection and anomaly investigation. It can be extremely challenging to repair a spacecraft once in orbit. That is why at ESA we focus on early detection of anomalies before they are serious and take preventive action. Data-driven diagnostics is useful to support the anomaly investigation process, contributing to effectively finding the cause of the anomaly.
Machine Learning applied to space operations has proven useful in prediction tasks ranging from thermal power consumption, temperature profiles, or even predicting when a spacecraft will cross the Earth radiation belts.
Reasoning
Autonomous and dynamic missions
Reasoning is a branch of Artificial Intelligence that concerns the process of organizing the activities required to achieve a desired objective, typically for execution by intelligent agents, subject to a set of constraints/rules organized in a model. The team is working on:
- Planning and Scheduling
- Resource Optimization
- Constraint programming
- Multi-agent systems
We apply these competences to develop an approach to exploit artificial intelligence and machine learning techniques to improve standard mission planning, and to provide autonomous and dynamic missions planning capabilities in support of current and future missions.
In “product oriented” activities we have developed solutions for specific mission problems. For example, in collaboration with the Flight Control Teams of various missions we addressed problems connected with data/command management for the spacecraft, ground station passes planning and optimization, payload activity allocation.
In “process driven” activities we develop general purpose tools for facilitating the design and synthesis of new products, exploiting the model based, domain independent nature of AI approaches. The general pursued idea is the one of improving the “process” of tool development, taking advantage of the state of the art AI planning and scheduling technology. Projects have been funded to deploy a platform for rapid prototyping of planning and scheduling applications (APSI) and for integrating it into standard mission planning activities (APSTR).
Interact
Highly customizable web based applications
Interact concerns all the topics related with the interaction of the cognitive system with the operational environment. It spans from the design of innovative interfaces between the human and the AI till the actual acting of robotics systems able to physically interact with the environment. On this topic the team is particularly active on:
- Data Visualization and Visual Analytics
- Plan Execution for Autonomous Systems
- Explainable AI
The need for accessing spacecraft mission data in a semantic, more profitable way has risen in recent years. In this context our group is working on highly customizable web based applications. Users do not to go through complex installation and update processes and they can access to their mission data from any device connected to the internet and able to run web browsers (laptops, tablets, smartphones...). We are committed to deliver solutions which rely on the combination of generic concepts with a very flexible configuration layer that leads to an effortless integration of future mission and applications.
Upcoming missions will require a higher degree of remote operations to increase quality and quantity of science return. Remote operations are certainly a challenging scenario, mainly because communication delays and errors lead to objective difficulties for the operator to promptly enable opportunistic science, activate contingency recovery procedures or to anticipate and react to tracked events.
AI planning-based control layers have demonstrated to be able to entail autonomy, and in this context we work on providing planning capabilities for on-board deliberation in support of autonomy for robotics and satellites.