FAQs
AI CHALLENGE COMPETITION
1. Eligibility and participation
Can an individual apply?
Yes. Eligible applicants include individuals such as students and researchers, as well as legal entities, SMEs/start-ups, universities, research organisations, NGOs, non-profits and consortia. Individuals must be at least 18 years old and meet the country and technical-capacity requirements.
Is a student living in Germany eligible if they have a different nationality?
The published rules frame individual eligibility by country of residence/establishment rather than nationality. Germany is an eligible EU Member State, and the application form asks for the applicant’s country of residence or registration. On that basis, a student residing in Germany appears eligible, provided all other requirements are met.
Is India an eligible country?
No. India is not included in the eligible EU Member States, Overseas Countries and Territories, or Horizon Europe Associated Countries listed in the competition documents.
Can an informal group of people or a team across different affiliations apply, or must they first create a legal entity?
A legal entity is not mandatory. The competition is open to one or more individuals, teams and consortia. The application form includes a field asking consortium applicants to explain their structure.
Must every member of a team or consortium be based in an eligible country, or is an eligible lead applicant sufficient?
All members inside the consortium are required to be resident/established in an eligible country.
Can the same organisation submit applications to more than one challenge through different internal teams?
Only one proposal per applicant will be evaluated. If the same applicant submits more than once, only the last submission received by F6S enters evaluation; all earlier submissions are declared non-eligible. Therefore, different teams applying under the same legal entity would normally be treated as one applicant. Legally distinct entities may each submit one proposal.
Are researchers affiliated with an organisation that is part of the AI-BOOST consortium eligible?
Cases of conflict of interest will be assessed by the AI Challenge Competition Office. Please share the specific case to the email account of the project: info@aiboost-project.eu.
Are participants expected to work on the challenge full-time?
No full-time or minimum full-time-equivalent requirement is stated. However, Advance Phase teams must actively participate for five months, develop and validate the solution, attend mandatory activities, submit required deliverables and demonstrate sufficient progress at the Mid-Term Checkpoint.
2. Application and SPARK Phase
What is the SPARK Phase deliverable? Does it require software development or a running simulation?
The SPARK Phase is the application and evaluation stage. Applicants submit a Concept Note through the F6S online form, describing the proposed solution, team capabilities and implementation plan. A running implementation is not required at submission, although applicants should provide credible evidence that the concept is technically feasible and can reach the challenge target during the Advance Phase.
Is there a separate Concept Note template?
The Concept Note is the F6S online application form. Annex 1 of the Guidelines reproduces the application questions so applicants can prepare their responses in advance. The online F6S version is the authoritative version.
Should applicants complete the Concept Note online, or upload a separate PDF application?
Applicants must complete and submit the official online F6S application form. A separate PDF is not accepted as a substitute for the online application. Applicants should only upload supporting files where the live F6S form explicitly requests them.
Does a consortium need to declare its internal budget or prize split in the application?
The published application form asks applicants to explain the consortium structure, work plan, team expertise and required resources, but it does not request a detailed consortium budget split. The Guidelines also do not specify how a prize should be distributed internally among consortium members. It’s up to the coordinator how the prize is split amongst the members of the consortium.
The challenge documents already provide KPIs. What additional KPIs should applicants include?
Applicants should use the challenge KPIs as the core technical benchmarks and translate them into proposal-specific, measurable outcomes. Additional KPIs may cover delivery milestones, processing time, resource efficiency, robustness, integration effort, scalability, environmental performance or user/industry impact. They should be realistic, quantified and directly linked to the proposed solution and work plan.
3. Competition phases, prizes and monitoring
Does “five SPARK winners per challenge” mean 20 SPARK winners in total?
Yes. There are four challenges and up to five SPARK winners per challenge, resulting in up to 20 SPARK winners. Each receives a EUR 28,500 SPARK prize, paid in two instalments subject to the competition conditions.
Will only four SMEs be funded in total, one per challenge?
No. The competition is not limited to SMEs, and the SPARK Phase may fund up to 20 eligible winners across all applicant types. At the end of the Advance Phase, one final winner per challenge – four in total – receives the EUR 100,000 final challenge prize. The winners may be individuals, research teams, universities, start-ups, SMEs or other eligible applicants.
What happens if a team does not pass the Mid-Term Checkpoint? Can it amend and resubmit, or is participation terminated?
The Mid-Term Checkpoint is pass/fail. Successful completion triggers payment of the remaining 50% of the SPARK prize. The Guidelines allow AI-BOOST consortium to withhold payments where requirements or active participation are not fulfilled.
4. Data, infrastructure, training and support
Do applicants receive challenge data or CINECA access during the SPARK Phase?
Generally, no. SPARK is the Concept Note stage. Detailed access to challenge datasets, technical resources and optional CINECA infrastructure is intended for teams selected for the Advance Phase and is introduced at the kick-off. Applicants may already use any openly available datasets listed in the relevant Challenge Description.
When will challenge data be shared, and is demo/sample data available before selection?
Challenge-specific access conditions are described in each Challenge Document. Selected teams receive detailed access information to the data at the Advance Phase kick-off. Before selection, applicants may use public datasets linked in the Challenge Documents.
Can applicants contact Challenge Owner experts directly during the SPARK Phase?
The published process provides a central AI-BOOST helpdesk for application questions. Direct, structured interaction with the Challenge Owner is explicitly planned for selected teams at the Advance Phase kick-off and through the whole Advance phase. No separate Challenge Owner contact points are published for SPARK applicants.
Where should technical questions be sent?
Technical and application questions should be sent to info@aiboost-project.eu.
When will training take place and what will it cover?
Specific training will take place to onboard selected applicants in the Applicants commit to participating in relevant AI-BOOST training and capacity-building activities. CINECA onboarding is specifically offered to selected Advance Phase teams that request HPC access; the detailed schedule is communicated directly to them. Topics may include account and environment setup, Python environments, SLURM job submission, data storage/transfer, AI frameworks, monitoring tools and distributed training/scaling methods.
When and how will Challenge Owner expert support be provided?
The main support period is the five-month Advance Phase. It starts with a kick-off and continues through technical guidance sessions, Q&A activities, webinars, regular feedback and the Mid-Term feedback session. Challenge-specific support includes, for example, robotics co-creation for Challenge 1, CAD/simulation mentoring for Challenge 2, clinical and secure-environment guidance for Challenge 3, and automotive validation mentorship for Challenge 4.
Will technical clarifications submitted by email be made available to all applicants?
Relevant clarifications should be shared consistently to preserve equal treatment, but the published documents do not specify a formal mechanism for publishing all email questions and answers.
5. Intellectual property, confidentiality and open source
Who owns the intellectual property created during the competition?
The participant retains ownership of the outcomes developed during the competition. The EU/awarding authority may use non-sensitive materials for policy, communication and dissemination purposes. In addition, the relevant Challenge Owner receives an exclusive, royalty-free right for six months after the competition to use, test, evaluate and implement the solution. Any later commercial exploitation, licensing or IP-sharing arrangement must be negotiated separately between the participant and Challenge Owner.
Can participants publish part of the solution as open source?
There is no general requirement to open-source the solution. Participants may propose open-source components or publications, and the application explicitly considers open-science contributions. This must remain compatible with third-party licences, confidentiality/NDA obligations, data restrictions and the Challenge Owner’s six-month exploitation right. Challenge 4 positively values open, shareable and reproducible tools; Challenge 2 proprietary data must not be disclosed.
6. Challenge 1 - Natural-language mission generation for agricultural robots
Is ROS1 mandatory? Can a ROS2 solution or ROS1-ROS2 bridge be used?
ROS-1 type is expected for Challenge 1. ROS2 model will not work; therefore is recommended to develop a model flexible enough to support both versions.
Must the solution use a VLA model, or can another Generative AI architecture be proposed?
The VLA solution is a requirement from the challenge owner, if the candidates presents something that does not fulfill the requirements in the evaluation phase they will be penalised.
Should the model output low-level continuous robot actions or high-level mission primitives?
The published scope points to high-level mission generation: speech/text is interpreted, decomposed into tasks and translated into structured ROS-compatible mission plans or command sequences. Evaluation compares generated task sequences with reference plans and successful execution. The documents do not require end-to-end low-level continuous control in the style of a generalist manipulation policy.
Must all model components be deployed directly on the physical robot, and which robotic platform will be used?
The final Proof of Concept must generate and execute robot missions and be validated on the provided robotic platform in a vineyard-row environment. The documents do not require every model component to run on-board, nor do they identify the exact robot model, sensors, compute hardware or software versions.
What Challenge 1 datasets and robot actions will be provided, and how large is the dataset?
The Challenge Owner will provide a domain-expert-labelled grape-cluster dataset, public vineyard/grape datasets, perception training data, and structured robot capability descriptions listing available actions and constraints. Real-time vineyard data will be used for final evaluation. The total number of images, exact class distribution and full action catalogue will be provided after the Advance Phase KoM.
Will a simulation environment, URDF, maps, point clouds, lidar scans or teleoperated demonstrations be provided?
The Challenge Description confirms access to a robotic platform, environmental maps, contextual data and a controlled vineyard testing environment. It does not commit to a specific simulator, robot URDF, lidar/point-cloud data, teleoperated demonstrations or action-labelled trajectories.
What physical tasks are expected, and is the interaction real-time or based on preconfigured routines?
The intended interaction is voice or text instruction translated into executable missions, with an optional target of 5-10 seconds from speech input to command generation. The system is assessed against a predefined set of tasks/actions and should support perception, navigation and task execution. The exact missions and whether they include physical manipulation, spraying, harvesting or only navigation/monitoring will be specified after the Advance Phase KoM.
7. Challenge 2 - Agentic AI for CAD generation and autonomous simulation
What file formats will be provided for the 2D drawings and 3D CAD models? Will 3D models be exportable to STEP?
The Challenge Description states that the core data consists of 2D piping drawings and 3D CAD piping models. The specific format of 3D models are available either in SolidWorks format (.sldprt / .sldasm) or in ESApro format, depending on whether they were designed before or after 2020. 2D models will be disclosed at the Advance Phase KoM. The models can be exported to STEP: SolidWorks supports STEP AP203/AP214 (.step / .stp), while ESApro also supports .step / .stp.
Where does the relevant engineering knowledge currently reside, and should the system accumulate knowledge across projects?
The source knowledge is described as being embedded in CAD assemblies, 2D drawings, engineering metadata, reviewed technical documentation and the practices of experienced engineers. The solution is expected to extract geometric, assembly and functional constraints from these sources.
Is the objective of Challenge 2 only to create a dataset?
No. The objective is to build a functional agentic AI Proof of Concept that interprets engineering requirements, extracts constraints, generates compatible parametric CAD/piping outputs, supports meshing and convergence analysis, and produces results suitable for expert review. The dataset is an input and validation resource, not the final output.
What would distinguish a top-scoring Challenge 2 proposal from a technically competent one?
A strong proposal must combine both domain credibility and AI innovation. It should demonstrate clear challenge fit, a feasible CAD/simulation workflow, novel agentic or geometric-reasoning methods, measurable reduction of engineering effort and package size, safe expert oversight, scalability, and a credible five-month implementation plan. The general SPARK assessment gives substantial weight to excellence/novelty, impact and quality of implementation; technical CAD competence alone or a generic AI concept alone would be insufficient.
8. Challenge 3 - Generative AI for enhancement of clinical datasets
How and when can participants access the Challenge 3 dataset? Is registration required before access?
The full clinical dataset is sensitive and can only be accessed and processed inside the Secure Processing Environment; it cannot be downloaded or copied to a personal HPC area. Detailed access arrangements are intended for selected Advance Phase teams and will be communicated at kick-off. Applicants should review and be prepared to accept the EUCAIM and dataset terms and conditions.
What is the expected TRL for Challenge 3: TRL 4-5 or TRL 7?
The general competition targets TRL 4-5, while the detailed Challenge 3 Description states TRL 7. The Guidelines explicitly say that the detailed Challenge Document prevails where there is a discrepancy. Therefore, the currently published challenge-specific target is TRL 7.
Does the requirement to balance underrepresented subgroups (as outlined in KPI 2) necessitate generating new, fully synthetic DICOM image series for newly created synthetic patient profiles?
Yes. To achieve the objective of KPI 2 (Bias Reduction and Cohort Balancing), participants are expected to generate new, fully synthetic patients that augment the original cohort and improve the representation of underrepresented demographic and clinical subgroups.
These synthetic patients should include both the clinical/tabular information and the corresponding CT imaging studies, forming complete synthetic patient records. The goal is to create synthetic cohorts that can be used to improve dataset balance while preserving the statistical and clinical characteristics of the original data.
This is distinct from the image generation task under KPI 3, where participants are required to generate missing CT acquisitions for existing patients after imaging series have been artificially masked for evaluation purposes.
Or is synthetic image generation strictly required for completing missing image modalities (e.g., Non-contrast CT, Contrast-enhanced CT, Low-dose CT) for existing patients as defined under KPI 3?
Yes. Under KPI 3, participants are expected to generate missing CT acquisitions for existing patients when one or more of the required CT series are unavailable during the evaluation. Missing imaging acquisitions will be artificially masked, and participants’ methods will be evaluated on their ability to reconstruct them according to the Imaging Completion Rate (ICR) metric.
Are all clinical variables listed in the JSON schema fully populated, or do missing values exist in the raw dataset? If missing, could you provide the general missingness ratios per variable?
For the evaluation of KPI 3, clinical variables will be artificially masked and participants will be required to reconstruct them using the original values as ground truth. We are not providing variable-specific missingness ratios, as participants will be able to determine these directly once they have access to the dataset within the challenge environment.
Does every subject possess all three CT series types (Non-contrast CT, Contrast-enhanced CT, and Low-dose CT)? If not:
- What is the availability distribution of these three series across the patient cohort?
The evaluation is based on artificially masking imaging acquisitions to assess image completion performance. We are not providing detailed statistics on the availability of each CT series or their distribution across subjects prior to the challenge, as this information will be available to participants once they have access to the dataset within the secure challenge environment. Participants should design methods capable of handling missing imaging acquisitions as defined in KPI 3.
- For patients with multiple CT series, are these series spatially paired/co-registered?
No. The CT acquisitions have not been spatially co-registered. The images correspond to the original clinical acquisitions and have not undergone any preprocessing after image acquisition and reconstruction on the CT scanner. Participants should therefore not assume voxel-wise alignment between different CT series belonging to the same subject.
What is the overall distribution of peak kilovoltage output (kVp) values across the dataset? Do any individual patients have multiple scans acquired at different kVp settings (whether paired or unpaired)?
The DICOM KVP tag (0018,0060) is one of the key acquisition parameters for the imaging harmonisation task, as it has a direct impact on CT image intensity distributions. The dataset contains heterogeneity in acquisition protocols, and different imaging studies may have been acquired using different acquisition parameters, including different kVp settings.
The objective of the harmonisation task is to reduce variability in CT image intensity distributions across scans acquired under heterogeneous acquisition protocols, with kVp being one of the main factors contributing to this variability. Participants are free to consider additional DICOM acquisition parameters if they believe these contribute to improving harmonisation performance.
We are not providing the distribution of kVp values across the dataset or patient-level statistics prior to the challenge since participants will be able to extract this information directly from the DICOM metadata once they have access to the dataset within the secure challenge environment.
9. Challenge 4 - Test-case generation from crash databases and standards
When and where will sample crash databases, reports and standards be available?
Challenge 4 primarily relies on public or otherwise lawfully accessible datasets. The Challenge Description lists representative open accident databases, multimodal datasets, OpenSCENARIO resources and Euro NCAP materials that applicants may access directly from the original providers. No single mandatory repository or separate sample-data release through F6S is promised. Where a source requires registration or approval, participants must obtain access independently and comply with its terms.