BOLERO on-Board cOntinual LEarning foR satcOm systems

  • Status
    Ongoing
  • Status date
    2025-07-22
  • Activity Code
    1B.138
Objectives

The objectives of BOLERO were multi-fold:

  • To identify potential satellite telecommunication (satcom) functionality and applications that could benefit from the use of continual learning methodologies in a fully transparent, quantifiable and reproducible way which will be reusable in future satcom and other missions for an informed selection of on-board machine learning (ML) models that could benefit from continual learning techniques.
  • To explore, develop, and simulate different continual learning implementation techniques for the identified satcom applications. This project objective also aimed to explore the connection between offline and online learning and the current state-of-the-art methodologies that would allow models that have been pre-trained offline to be updated so they can be enhanced by online/continual learning.
  • To identify and justify a suitable system architecture for on-board continual learning applications through performing the benchmarking process of all developed ML algorithms with continual learning techniques for all satcom applications and simulation scenarios, as well as through performing the theoretical trade-off analysis of the hardware and system architectures considered in BOLERO.
Challenges

The most important challenges of BOLERO related to:

  • Availability of hardware architectures that could be used for benchmarking continual learning AI algorithms.
  • Availability of datasets that could be used to verify and validate continual learning scenarios (therefore, we developed synthetic data simulators for all selected satellite communications applications).
  • Building reproducible, unbiased and reproducible pipelines for the quantitative validation of AI algorithms.
  • Developing an objective procedure for selecting satcom use-cases that would benefit from continual learning paradigms.
Benefits

The technology solutions develop in BOLERO bring important benefits that are transferable to the current and future (not only) satcom missions – these potential benefits include:

  • The possibility of synthesising the datasets in selected satcom applications (anomaly detection in telemetry data, congestion prediction with flexible payload, and inter-satellite link optimisation);
  • Ready-to-use and thoroughly validated continual learning algorithms for satcom applications that can adapt their behavior to the changing data characteristics;
  • The possibility of performing fully objective, quantifiable and reproducible analysis of (not only) satcom applications for on-board deployment in continual learning settings, and of hardware architectures that can be considered for on-board deployment in continual learning settings;
  • The possibility of understanding the most pressing research and development gaps through analysing the developed research and development roadmaps that shall be followed to accelerate the adoption of continual learning in real-world satcom missions.
Features

The BOLERO technology is composed of several pivotal components, including:

  • An assessment matrix that can be used to objectively select the most appropriate satcom applications for on-board continual learning implementation (i.e., the use-cases and on-board applications that could benefit most from continual learning);
  • An assessment matrix that can be used to quantify the applicability of the analysed hardware architectures in on-board implementation, with a special emphasis put on on-board continual learning algorithms;
  • Synthetic data generators, developed for each considered satcom application (anomaly detection in satellite telemetry, beam-hopping, and inter-satellite routing) that can be used to synthetically generate data for simulated continual learning scenarios in a fully reproducible and traceable way;
  • Continual learning AI algorithms for the selected satcom applications, dealing with the catastrophic forgetting phenomenon and addressing different continual learning strategies (including class-incremental learning, task-incremental learning, and domain-incremental learning).

The roadmaps, presenting the most important activities that need to be followed to accelerate the adoption of continual learning technologies in satcom systems. These roadmaps have been split into those that relate to the algorithms, technologies, hardware as well as programmes, the latter indicating the programmatic gaps that were identified in this activity.

System Architecture

The technology developed in BOLERO is fully modular, and directly relates to the key product features, including:

  • Synthetic data generators;
  • Continual learning artificial intelligence algorithms developed for selected satcom applications;
  • Assessment matrices for selecting a) appropriate satcom applications for on-board continual learning deployment and the b) best hardware architectures for such on-board implementation;
  • Research and development roadmaps. All these components are stand-alone and self-contained entities that can be effectively used separately.
Plan

Project was planned and divided into specific Work Packages focusing on the following:

  • WP100 SOTA: Review and analysis;
  • WP200 Satcom applications: Identification and analysis;
  • WP300 Algorithms: Continual learning for satcom;
  • WP400 Hardware: Performance assessment;
  • WP500 Programmatic and development gaps;
  • WP600 Software;
  • WP700 Management, outreach and dissemination.
Current status

In BOLERO, we identified satcom functionalities that could benefit from on-board continual learning, selected them via a quantifiable process, and developed ML models accordingly.

We built data simulators to test various continual learning techniques, emphasizing reproducibility and algorithm generalization in realistic settings. The project delivered an end-to-end pipeline for continual learning and online adaptation for satcom, validated in simulated scenarios. Finally, we benchmarked these methods across different hardware, including KP Labs’ Leopard and BrainChip’s Akida, providing comprehensive results for all algorithms, hardware, and applications explored within BOLERO.

Prime Contractor

Subcontractors