The Combat Air AI Challenge is a joint endeavour with ministry of Defence (MOD) Defence Science and Technology Laboratory (Dstl) and is closely related to Leonardo UK’s work within GCAP, which is to develop and deploy Integrated Sensing And Non-Kinetic Effects (ISANKE) and Integrated Communication Systems (ICS). The ISANKE super-system is a complex network of capabilities that sits across the combat air domain and, whilst co-ordinating numerous sensors, assimilates information into a common situational awareness picture as part of an integrated offering.
ISANKE will unlock the potential of sixth generation tactical sensing by transitioning individual airborne sensors towards a fully integrated sensing, fusion and self-protection capability. ICS will enable ISANKE to operate as an adaptive mesh network across formations of crewed and un-crewed aircraft. Collectively, ISANKE and ICS will have the capability to dynamically self-optimise how sensing and effecting are best performed across the formation in a given tactical situation. Ultimately, the ISANKE and ICS capability will help the aircrew to survive and operate more effectively in the complex and highly contested battlespace of tomorrow.
The Combat Air AI Challenge consists of two streams:
-
Stream 1 looks to utilise Deep Reinforcement Learning to inform sensor resource management across the ISANKE super-system by modelling behaviour of the Blue Force platform whilst exposed to a range of operational conditions.
-
Stream 2 intends to explore potential deployment of deep integrated sensor fusion across the ISANKE super-system using Machine Learning.
Each stream will consist of a pilot project with nominated participants, lasting approximately six months in duration and requiring completion by December 2023. We anticipate that the successful completion of pilot projects will potentially lead to wider collaboration(s) in the GCAP ISANKE problem space.
The Combat Air AI Challenge intends to act as a high profile, fast-paced, open innovation incubator to test out disruptive ideas and potential longer-term collaboration opportunities in the GCAP domain. In partnership with Dstl, as part of Team Tempest, £250,000 in total has been provisioned to support nominated participants develop their solution.
By drawing on the state-of-the-art knowledge-base of start-ups, SMEs and academia, we hope to unlock the impact of AI within GCAP. This will also help build a UK eco-system of skilled professionals, as well as organisations adept at operating within this problem space.
You can find the problem statements for Stream 1 & 2, and more details on how to respond, in the tabs below.
Stream 1 of the Leonardo Combat Air AI Challenge intends to explore potential deployment of deep reinforcement learning across the ISANKE super-system by modelling behaviour of the Blue Force platform whilst exposed to a range of operational conditions.
The challenge is to demonstrate the ability to create an algorithm that is capable of learning to play a simplified air combat game called “2042”. Although highly stylised, the game captures a number of salient aspects of modern air combat.
There will be opportunity to discuss the finer technical aspects of this stream upon commencement of the pilot project whereby, the game engine will also be released to the nominated participants.
You will be given:
-
The “2042” game engine (consisting of roughly 2000 lines of Python Code)
-
Configuration files for a number scenarios
-
Instructions on which parts of the game engine may be accessed to provide data for your learning algorithm
-
Another program (around 600 lines of Python) that can be used to visualise and interact with the scenario, e.g. to play against the AI you have developed
You will need to:
-
Develop a Deep Reinforcement Learning algorithm and interface the code with the game engine
-
Train your model on the scenarios starting from the simplest – available likely scenarios are:
i. Elementary flying
-
Aircraft can fly over the objective, return to base and land
ii. Learn to interpret tracks and fire weapons
- 1 vs 1…n
iii. Learn to implement operational constraints, evade enemy missiles
- 1 vs 1…n
iv. Learn to passively co-operate with friendlies using the network
- 2…n vs 1…n
v. Play the AI against itself
- 1…n vs 1…n
Expected outputs from the Stream 1 Pilot Project:
-
The developed algorithm including source code, the script that trains the model and a data file containing learned model parameters
-
Report detailing the rationale behind the technical approach adopted
-
Demonstration of the software playing the 2042 game
Success Criteria:
-
Number of “wins” in the game by the algorithm developed
-
Run-time and processing capability required for training the algorithm
-
Run-time and processing capability required for the algorithm to play the game
-
Ability to adapt the algorithm to changes in the configuration and/or new actions that could be added
Stream 2 of the Leonardo Combat Air AI Challenge intends to explore potential deployment of deep integrated multi-platform sensor fusion across the ISANKE super-system using Machine Learning.
There will be opportunity to discuss the finer technical aspects of this stream upon commencement of the pilot project whereby, the below defined training data-set will also be released to the nominated participants.
You will be given:
-
Imagery of five types of target (e.g. different types of Red Force platforms)
-
IR imagery and EO imagery which is such that each target subtends approximately 20 pixels and each image is 100 x 100 pixels or thereabouts
-
Approximately 500 examples of images of each target seen by two sensors from two platforms (i.e. approx. 500 x 2 x 2 = 2000 images per target)
-
Ground-truth on:
- Position of target in the images to assess “detection”/localisation-in-image performance.
- Target type to assess classification performance.
- Viewing angle and range to assess how these influence the classifier performance.
The training data-set above would be representative of imagery generated from multiple Blue Force platforms observing a distant object that potentially could turn out to be a Red Force platform.
This data-set will be provided by our existing collaborators, the University of Liverpool, through their simulation capability.
You will need to:
-
Develop a machine learning algorithm for “multi-view fusion”. In multi-view fusion, multiple views of a target image are input into a single classifier and the classification decision is learned as a function of the combined input
-
Implement an alternative “single-view” identification approach where the outputs from classifiers trained to process single images are combined, for example by multiplying the four classifier scores for each class from various classifiers
-
Compare the two classifiers, one for each of 1. and 2., using the training data and assess performance of both approaches on a held-out validation data-set (to be provided towards final stages of the pilot project and drawn from the same set as the training data)
Expected outputs from the Stream 2 Pilot Project:
-
Source code for all classifiers
-
A concluding demonstration of the results and performance evaluation on held-out validation data-set
-
A report detailing technical development approaches
Success Criteria:
-
Improvement in classification accuracy on a validation data-set (assumed to comprise 20% of the data that is initially provided) offered by “multi-view fusion” relative to fusion of “single-view” classification outputs
-
Classification accuracies (of both approaches) achieved on the validation data-set
-
Run-time and processing capability required for training of the algorithms (on a training data-set assumed to comprise 80% of the data that is initially provided)
-
Run-time and processing capability required for applying the trained algorithms to one set of four (previously unseen) images
Your organisation details:
-
Registered organisation name, UK registration number and address
-
A Point of Contact for communication purposes
-
Tell us a bit more about your organisation e.g. size, primary capabilities and anything else that could support your proposal
-
Have you previously worked with Leonardo? This is for us to assess timescales required for NDAs etc as/where applicable
-
By responding to this challenge, you confirm that, should your organisation be nominated for pilot project(s), you have deployable resources to facilitate completion by 31 Dec 2023
Outline Proposal:
-
Which stream(s) of the challenge you are responding to.
-
Outline of proposed technical approaches and the rationale behind those to tackle your chosen stream(s).
-
A workplan detailing completion by 31 Dec 2023
-
A vision outlining how your proposed approaches can be exploited into game-changing UK Combat Air capability by 2035 beyond the scope of the pilot project(s).
Questions?
Contact us at map.aiml23@leonardo.com
Leonardo is committed to ensuring the security and protection of the personal information that we process, and to provide a compliant and consistent approach to data protection in accordance with the Data Protection Act 2018.
If you have any questions about the challenge, please contact us via email at map.aiml23@leonardo.com
Please note that this email is used solely for the purpose of the Combat Air AI Challenge related communications. We are no longer accepting proposals.
Leonardo is committed to ensuring the security and protection of the personal information that we process, and to provide a compliant and consistent approach to data protection in accordance with the Data Protection Act 2018.