EIC Advanced Innovation Challenges Results 2026: 20 Selected Projects

The European Innovation Council has selected the first 20 EIC Advanced Innovation Challenges projects: 10 Physical AI projects and 10 New Approach Methodologies projects. Together, they form a €6 million Stage 1 portfolio and will compete for up to €2.5 million per project in Stage 2.

The cohort is the first practical test of a new challenge-driven, stage-gated EIC funding model inspired by ARPA-style programmes. It is not another EIC Accelerator results round. The 20 projects started on 1 September 2026 and have only nine months to generate comparative evidence, validate a demanding use case and turn initial stakeholder interest into a credible route toward deployment.

Competition was severe. The call received 709 proposals from 39 countries, but only 20 were selected. That is an overall success rate of 2.82%, or approximately one selected project for every 35 applications. The selected portfolio reveals the EIC's priorities and the evidence each team must produce to remain competitive for Stage 2.

EIC Advanced Innovation Challenges 2026 Results at a Glance

• Results announced: 10 September 2026
• Proposals submitted: 709 from 39 countries
• Physical AI proposals: 425
• New Approach Methodologies proposals: 284
• Projects selected: 20, divided equally between the two Challenges
• Countries represented by selected coordinators: 10
• Overall proposal success rate: 2.82%
• Stage 1 grant: €300,000 as a fixed lump sum for each project
• Total Stage 1 portfolio: €6 million
• Stage 1 period: 1 September 2026 to 31 May 2027
• Stage 2 opportunity: up to €2.5 million per selected project for up to 2.5 years
• Current indicative Stage 2 deadline: 18 June 2027
• Stage 2 access: restricted to projects selected and funded in Stage 1

Contents

1. What the EIC Advanced Innovation Challenges pilot is
2. Selection rates and oversubscription
3. How Stage 1 and Stage 2 funding work
4. The 10 selected Physical AI projects
5. What the Physical AI portfolio reveals
6. The 10 selected New Approach Methodologies projects
7. What the NAM portfolio reveals
8. Country distribution across the 20 projects
9. What must happen before Stage 2
10. Lessons for applicants, users and investors
11. Frequently asked questions
12. Sources and methodology

What Is the EIC Advanced Innovation Challenges Pilot?

The EIC Advanced Innovation Challenges, commonly shortened to EIC AIC, are a Horizon Europe pilot for high-risk, demand-driven deep tech innovation. The scheme targets fields in which Europe has substantial research capacity but has struggled to convert that knowledge into products, operational systems and broad market adoption. Its defining feature is not simply the deep tech subject matter. It is the combination of a narrowly defined challenge, short competitive stages, measurable milestones, direct user involvement and portfolio-level management by an EIC Programme Manager.

The pilot therefore begins from a problem and a desired operational outcome rather than from an unrestricted technology category. The 2026 call defined two problems: how to move Physical AI systems into complex real environments, and how to move human-relevant New Approach Methodologies into biomedical and regulatory practice. Applicants had to arrive with technology at approximately Technology Readiness Level 4, access to suitable testing infrastructure and documented interest from an end user, integrator, industrial stakeholder or regulator.

The programme is described as ARPA-style because it borrows several ideas associated with mission-led advanced research programmes: empowered programme management, defined technical challenges, active portfolio steering, rapid experimentation, milestone-based continuation and competition between multiple approaches to the same strategic problem. It remains an EIC grant scheme operating under the Horizon Europe legal framework; it is not a direct copy of the United States Defense Advanced Research Projects Agency.

How AIC Differs From the EIC Accelerator

A university, research institute or foundation can lead an AIC Stage 1 project, and the €300,000 award does not represent an investment decision. Company financing selections are reported separately in the aggregated EIC Accelerator results.

Selection Rates: 20 Projects From 709 Proposals

The EIC reported 709 submissions requesting a combined €130.7 million. The €6 million Stage 1 budget was therefore oversubscribed by approximately 21.8 times when measured against requested funding. Proposal demand was not evenly divided: Physical AI attracted 425 proposals, while the NAM Challenge attracted 284. The final portfolio nevertheless reserved 10 positions for each Challenge.

A Physical AI proposal faced a selection rate approximately one-third lower than a NAM proposal because 141 more applications competed for the same number of positions. That does not show that the Physical AI evaluations were harsher. It shows the mathematical effect of an equal 10-project allocation across two applicant pools of different sizes.

The applicant population was itself commercially weighted. The EIC counted 490 company applicants, representing 69% of the total, followed by 124 higher-education establishments and 89 research organisations. Germany generated the most proposals with 84, followed by Italy with 71 and the Netherlands with 62. All three countries also appear repeatedly in the final cohort, although the published aggregate data are not sufficient to calculate comparable success rates for every participating country.

Why the 2.82% Rate Needs Context

The 2.82% figure measures entry into Stage 1, not completion of the full programme. These 20 projects now compete inside the portfolio. Stage 1 funding pays them to produce the evidence on which the next selection will be based. The final conversion rate from 709 original proposals to Stage 2 awards will therefore be lower and will not be known until the 2027 restricted call is evaluated.

Portfolio selection also means that rank order alone does not explain the result. Proposals first had to clear the quality thresholds, after which the EIC could choose a combination of approaches, technologies, use cases and risk profiles that collectively covered each Challenge. A technically excellent project could consequently remain unfunded if it duplicated another approach or left the overall portfolio less balanced.

How the Two-Stage Funding Model Works

Every selected project has the same Stage 1 contribution and the same recorded project dates. The grant is a fixed lump sum, not an EIC Fund investment and not a reimbursement ceiling under which a beneficiary simply invoices every expense. Work packages, outputs and completion conditions govern payment under the Horizon Europe lump-sum model.

Stage 2 is competitive rather than automatic. The EIC's current programme page gives an indicative deadline of 18 June 2027, less than three weeks after the Stage 1 projects are scheduled to end. The 2026 Work Programme anticipates a €25 million Stage 2 budget, but expressly makes the 2027 phase subject to adoption under the 2027 Work Programme. If every Stage 2 project requested the €2.5 million maximum, €25 million would fund no more than 10 projects. That is an arithmetic ceiling, not a published promise that exactly 10 will be selected.

A project that passes both gates can receive up to €2.8 million across the two stages. The larger strategic value, however, is continuity: Stage 1 produces the data package, user relationships and regulatory or deployment plan needed to justify the much larger Stage 2 experiment.

Challenge 1: The 10 Selected Physical AI Projects

The official title of the first Challenge is Accelerating Physical AI: Embodied Intelligence for the Next Frontier of AI-Powered Robotics. Physical AI refers here to intelligent systems that perceive, reason, decide and act through physical machines in complex environments. Eligible concepts had to demonstrate at least two of five characteristics: intelligent perception and cognition; adaptive learning and optimisation; autonomous decision-making or collective intelligence; human-AI collaboration; and physical integration involving sensors, actuators or materials.

The call concentrated on three application areas: disaster response and civil security, autonomous laboratories for scientific discovery, and personal or professional robot assistants. The selected portfolio interprets those categories broadly, spanning warehouses, coasts, factories, energy infrastructure, laboratories, rescue operations and private homes.

1. ARAL: AI-Driven Robotic Assistant for Logistics

Grant ID 101329116; PAL Robotics SLU, Spain. ARAL starts from PAL Robotics' bipedal Kangaroo platform and targets varied work in high-throughput logistics centres. The system is intended to navigate dense, human-centred facilities, interact with existing shelving and equipment, and manipulate heterogeneous items ranging from deformable textiles and hanging garments to cartons and containers. That is a harder problem than moving standard boxes through a purpose-built automated warehouse because the robot must cope with variation, task switching and proximity to people.

ARAL combines perception and cognition, adaptive learning and compliant manipulation. Its Stage 1 proof point is not simply that a humanoid can perform a warehouse demonstration. It must generate comparative performance evidence showing that one embodied system can execute useful tasks reliably in legacy facilities without requiring a costly redesign of the surrounding infrastructure. The public CORDIS record also identifies interest from major retail end users, giving the project a concrete route for refining operational requirements.

2. Angler AI Agent: A Hybrid Maritime Physical AI System

Grant ID 101328431; Angler s. r. o., Czechia. The Angler AI agent combines an uncrewed surface vessel and an underwater vehicle in one autonomous platform for coastal monitoring and critical-infrastructure inspection. Its sensing package brings together sonar, LiDAR, optical and thermal data, while its autonomy layer must maintain navigation and mission control in environments where satellite positioning is degraded or unavailable.

The project's differentiating challenge is the transition between surface and subsea operation. Ports, subsea cables, offshore renewable-energy installations and coastal ecosystems are normally inspected through separate assets, crews and workflows. Angler is testing whether a single embodied agent can reduce that fragmentation while identifying anomalies over long missions. Stage 1 therefore has to establish credible autonomy, sensor fusion and cross-domain performance, not merely demonstrate that each vehicle works independently.

3. FLEXAMP: Autonomous Assembly of Flexible Wiring Harnesses

Grant ID 101327983; Fraunhofer-Gesellschaft, Germany. FLEXAMP addresses a manufacturing task that remains resistant to conventional automation: manipulating and assembling deformable cables and wiring harnesses. Rigid parts follow predictable geometry. Cables bend, twist, overlap and change shape as they are handled, making perception and control substantially more difficult.

The project uses a bimanual robot, teleoperation and multimodal visual, tactile and force data to learn reusable manipulation policies. The core commercial question is transferability. A policy that works only for one cable, connector or workstation has limited industrial value. FLEXAMP must show that its learned behaviours can remain dependable across changes in cable properties, task conditions and factory setups, reducing the reprogramming burden that has kept wiring-harness production labour-intensive.

4. ADAPT-AB: Autonomous Decision-Making for Antibiotic Discovery

Grant ID 101328938; Resistell AG, Switzerland. ADAPT-AB applies Physical AI to an autonomous laboratory rather than a mobile robot. It combines high-frequency nanomotion sensing, edge-AI interpretation, Resistell's ResiDaq platform and deterministic laboratory control. The objective is to turn automated equipment from a script executor into a closed-loop experimental system that can interpret biological responses while an experiment is running and adjust subsequent actions within predefined safety constraints.

Its initial application is antibiotic discovery and early in-vitro toxicology. The project must demonstrate that real-time sensing and decision-making shorten the experiment-learn-adjust cycle without sacrificing traceability or human oversight. That makes ADAPT-AB a test of scientific autonomy: the system needs to connect perception, interpretation and physical laboratory action in a way that produces faster and more useful evidence than post-experiment analysis alone.

5. PHYS-EXO: Physical Embodied Intelligence for Smart Exoskeletons

Grant ID 101326967; University of Applied Sciences and Arts of Southern Switzerland (SUPSI), Switzerland. PHYS-EXO seeks to move wearable exoskeletons beyond reactive assistance. Conventional systems primarily respond to posture, force or motion. PHYS-EXO adds egocentric sensing, semantic mapping, risk reasoning and task planning so that the wearable system can interpret the environment and provide proactive, context-aware assistance.

The planned use cases are urban search and rescue and industrial material handling. Both require a careful balance between useful autonomy and human control: an exoskeleton is attached to its user, so an incorrect decision has immediate physical consequences. The project therefore has to validate not only reduced strain or improved task performance, but also understandable guidance, reliable risk detection, low-latency edge operation and safe behaviour in changing surroundings.

6. FORMOVE: A Data Foundation for Physical AI

Grant ID 101326990; Formove GmbH, Germany. FORMOVE is the portfolio's clearest infrastructure play. Rather than building one end-use robot, it is developing a robot-agnostic pipeline for capturing, labelling and transforming real human motion and interaction into robot-training data. Its Human2Robot retargeting layer is designed to translate demonstrations across different robotic bodies.

Physical AI is constrained by the cost and scarcity of high-quality interaction data. Data collected for one robot, embodiment or task often transfers poorly to another. FORMOVE must show that its acquisition and retargeting process can create useful policies more efficiently and improve success on real manipulation tasks. If validated, its value would extend across personal and professional robotics because it supplies a reusable learning foundation rather than a single application.

7. ARIM: Autonomous Robotic Inspection and Manipulation

Grant ID 101328770; Energy Robotics GmbH, Germany. ARIM targets the transition from robots that observe industrial infrastructure to robots that physically intervene. Energy Robotics already operates in autonomous inspection; this project adds manipulation through a hardware-independent Physical AI stack built around perception, reasoning and hybrid vision-language-action models.

The distinction between identifying a defect and safely acting on equipment is substantial. Industrial sites contain unfamiliar geometries, changing conditions and high-consequence hazards. ARIM must benchmark whether its system can choose and execute useful interventions reliably across robotic platforms while preserving safety. Its strategic case combines dangerous maintenance work, shortages of skilled industrial labour and the need for more autonomous operation of European energy and critical infrastructure.

8. Robody: Humanoid Robots for Elderly Home Care

Grant ID 101329100; Devanthro GmbH, Germany. Robody develops a humanoid home-care assistant that combines remote human operation with gradually learned autonomy. Teleoperation supplies immediate human capability and produces training data. An uncertainty monitor is intended to return control to a human operator whenever the autonomous policy is not sufficiently confident.

The home is one of the least structured environments in robotics, and care adds exceptional safety, privacy and trust requirements. The proposed service spans monitoring, medication and chronic-disease support, physical assistance and activity. Stage 1 must establish that the human-in-the-loop architecture can deliver dependable help while creating a credible path toward one operator supervising several robots. The decisive evidence will concern safe task completion and escalation, not humanoid appearance.

9. APOLLO: An Agentic Physical AI Operating System for Self-Driving Laboratories

Grant ID 101328728; VIMI Labs UG, Germany. APOLLO is building an operating layer that connects AI agents, digital twins and laboratory robots. Digital agents first train and test robot behaviour in simulated environments, then orchestrate physical experiments under a guardian-controlled human-in-the-loop framework.

The validation cases focus on ink optimisation for batteries and fuel cells and on precision encapsulation for aerospace and energy components. APOLLO's challenge is interoperability: autonomous discovery remains difficult when instruments, robots, models and data systems operate as isolated components. Stage 1 must show that the operating system can coordinate those components safely and shorten real experimental cycles, rather than functioning only as a digital workflow demonstration.

10. HIVE: A High-Interoperability Environment for Humanoid Robots

Grant ID 101327261; Oversonic Robotics SRL Società Benefit, Italy. HIVE adds greater cognitive autonomy to Oversonic's RoBee industrial humanoid. Its onboard architecture combines vision, depth, LiDAR and force sensing with vision-language and vision-language-action models. On-premise processing is intended to reduce latency and support cybersecurity and data-sovereignty requirements.

The main demonstration concerns cleaning burn-in boards at an STMicroelectronics semiconductor facility. RoBee must handle boards, machines and ovens in a shared industrial workspace without rigid task-by-task programming. That use case gives HIVE an exact test environment and a demanding customer context. The project must prove repeatable task performance, safe collaboration and adaptation under factory conditions while showing that its architecture can transfer beyond one semiconductor process.

What the Physical AI Portfolio Reveals

Physical AI Is Broader Than Humanoid Robotics

Humanoids are prominent through ARAL, Robody and HIVE, but the portfolio does not reduce Physical AI to humanoid form. It includes a maritime vehicle, an exoskeleton, a bimanual assembly system, an autonomous laboratory control loop, an operating system, a training-data pipeline and an industrial manipulation stack. The shared requirement is closed-loop intelligence connected to physical action.

The EIC Selected Several Layers of the Same Emerging Stack

• Data generation and transfer: FORMOVE creates interaction data and retargeting methods.
• Policy and control: FLEXAMP, ARIM and PHYS-EXO develop adaptive behaviour for difficult physical tasks.
• System orchestration: APOLLO coordinates agents, simulations and laboratory equipment.
• Integrated robotic platforms: ARAL, Angler, Robody and HIVE test complete systems in defined environments.
• Closed-loop scientific operation: ADAPT-AB connects biological sensing directly to experimental decisions.

This layered composition is consistent with portfolio logic. Selecting only end-use robots would leave common bottlenecks in data, interoperability and control untouched. Selecting infrastructure alone would produce no demanding deployments. The chosen mix creates opportunities for cross-project learning while retaining competition between approaches.

End-User Environments Are Part of the Technology

Every project is tied to a setting in which failure modes can be measured: a warehouse, coastline, wiring workstation, biological laboratory, rescue scene, industrial facility, home or semiconductor plant. The environment is not an illustrative use case added after the AI system is built. It defines the sensors, response times, safety constraints, task variability and adoption barriers against which the system must be benchmarked.

Europe's Strategic-Autonomy Objective Is Visible

Several project summaries explicitly connect their technology to European autonomy: a European operating system for self-driving labs, on-premise industrial humanoid intelligence, robot-agnostic training data and autonomous inspection of critical infrastructure. The Work Programme names reduced dependence on non-European technology providers as an expected impact. This portfolio turns that broad policy goal into specific control points in the Physical AI value chain.

Challenge 2: The 10 Selected New Approach Methodologies Projects

The second Challenge is Translating Disruptive New Approach Methodologies into Practice. New Approach Methodologies, or NAMs, are methods intended to replace, reduce or refine animal use while improving the human relevance of biomedical research and product testing. The EIC notes that there is no single official definition for NAM in this call. Its scope includes organoids, organ-on-chip and disease-on-chip systems, in-chemico methods, digital twins, virtual patients, AI-enhanced predictive models, integrated in-silico platforms and advanced 3D human tissues.

The Challenge covers preclinical biomedical research and the safety, efficacy or quality testing of medicines and medical technologies. A scientifically interesting model was not enough. Applicants needed a TRL 4 solution, testing infrastructure and a letter of intent from an industrial end user or regulatory body. The selected projects consequently pair a technical method with a defined adoption problem.

11. STRIALIS: Mechanistic Subtyping for Parkinson's Disease

Grant ID 101328863; VIB VZW, Belgium. STRIALIS addresses a central weakness in Parkinson's disease development: patients grouped under the same clinical diagnosis can have different molecular causes, while therapies that look promising in conventional preclinical models repeatedly fail in late-stage trials. The project is building a human-relevant model and an initial atlas of mechanistically distinct disease subtypes.

The intended value is treatment matching. Instead of asking whether one candidate works in a generic Parkinson's model, STRIALIS asks which molecularly defined patient group is most likely to respond. Stage 1 must show that the platform can distinguish meaningful subtypes, support reproducible drug testing and improve the evidence used to advance or discard candidates. Regulatory preparation and assay throughput are essential because a precise model that cannot be standardised will not change industrial decisions.

12. STEM-SAFE: Multi-Organ Safety Testing for Stem-Cell Gene Therapies

Grant ID 101328580; Fondazione Telethon ETS, Italy. STEM-SAFE is validating MOAB-BM-LN, a dual-organ in-vitro system representing important features of human bone marrow and lymph nodes. It focuses on the tumour-forming risk of gene therapies based on haematopoietic stem and progenitor cells, including the detection of pre-leukaemic events using the actual therapy product.

The project sits at the intersection of scientific prediction and product safety. Simplified cell assays can miss multi-organ behaviour, while mouse models do not reproduce every human response. STEM-SAFE must demonstrate that its model produces sensitive, reproducible and decision-relevant safety evidence. Its longer-term ambition is not merely publication: it is an industry-standard and regulator-qualified method capable of supporting the development of advanced therapies.

13. AI4BirthCare: Patient-Specific Decision Support for Childbirth

Grant ID 101327625; INEGI, Portugal. AI4BirthCare combines information gathered during routine prenatal care with biomechanical simulation and machine learning. In the final weeks of pregnancy, the system is intended to estimate patient-specific risks, identify pelvic regions vulnerable to injury and support clinical decisions concerning delivery and follow-up.

This is a hybrid NAM: the relevant model is computational, but its utility depends on clinical data, biomechanical validity and integration into hospital practice. Stage 1 must establish predictive performance and demonstrate that the output can inform safer choices rather than simply produce another risk score. Data governance, clinical validation and understandable decision support are therefore as important as model accuracy.

14. VANTAGE: Vascularised Tumour-on-Chip Testing for Antibody-Drug Conjugates

Grant ID 101327116; University of Bern, Switzerland. VANTAGE is industrialising a perfused, human vascularised tumour-on-chip for evaluating antibody-drug conjugates. The platform is designed to measure primary tumour destruction, inhibition of metastatic seeding, off-target vascular toxicity and barrier function in breast-cancer and non-small-cell lung-cancer models.

Its central proposition is a broader efficacy-and-safety picture than conventional models provide. Vascular behaviour is critical to how an antibody-drug conjugate reaches a tumour and how toxicity manifests outside it. VANTAGE combines an injection-moulded platform with AI-based image metrics to pursue higher throughput and reproducibility. Stage 1 must prove that the readouts are stable, discriminating and relevant enough to alter development decisions.

15. InSteps: In-Silico Testing for Stroke Thrombectomy Devices

Grant ID 101328131; InSteps B.V., Netherlands. InSteps is developing a virtual testing platform for mechanical thrombectomy devices used to remove clots during stroke treatment. It combines finite-element simulation, generative AI and data from 4,500 patients to create diverse virtual cohorts and model the interaction between a device, vascular anatomy and human-derived blood clots.

Medical-device development currently depends heavily on iterative physical testing and animal models before clinical use. InSteps aims to reject weak designs earlier, explore anatomical variation at scale and improve the evidence supporting later regulatory and clinical work. Stage 1 must establish verification and predictive credibility: a visually sophisticated simulation has little decision value unless its behaviour can be traced to real device and patient outcomes.

16. SCALE-MYO: Scalable Engineered Human Muscle Tissues

Grant ID 101327207; University of Turin, Italy. SCALE-MYO targets a supply bottleneck in metabolic and neuromuscular drug discovery: reproducible, functional human skeletal-muscle tissue at industrial scale. Its GERALT genetic-programming platform is designed to remove expensive and variable growth-factor dependencies, allowing induced pluripotent stem cells to expand and differentiate more consistently.

The intended output is a standardised thaw-and-cast input that produces measurable force and metabolic behaviour on a practical timetable. SCALE-MYO must therefore validate manufacturing repeatability, tissue function and compatibility with microphysiological assays. The project's importance lies in treating scale as part of scientific validity. A tissue model cannot become a routine NAM if each laboratory batch behaves differently or requires specialist craft to produce.

17. OCENTRA: Automated Patient-Derived Organoid Screening

Grant ID 101329292; OKOMERA, France. OCENTRA integrates microfluidic culture, compound exposure, workflow traceability and AI analytics in a compact benchtop system. It is designed to convert small patient-derived tumour samples into standardised 3D organoids within 48 hours for target validation, dose-response studies and precision-medicine applications.

Patient-derived models can preserve clinically important tumour characteristics, but manual production and analysis restrict reproducibility and throughput. OCENTRA's Stage 1 test is whether automation can make those models dependable enough for pharmaceutical workflows without erasing their biological relevance. Independent validation in an industry setting, data governance and a credible regulatory route are central to proving that the platform can move beyond a specialised research service.

18. TEVAR-Twin: Digital-Twin Planning for Thoracic Aortic Repair

Grant ID 101328360; AllStent S.r.l., Italy. TEVAR-Twin creates a patient-specific three-dimensional model of the thoracic aorta from computed-tomography scans and uses finite-element simulation to test virtual deployment of commercially available stent grafts. Clinicians can compare intervention strategies and inspect predicted post-deployment geometry before performing thoracic endovascular aortic repair.

The method targets a planning process that remains dependent on manual measurements and clinician experience. A useful digital twin must do more than reconstruct anatomy: it must forecast device-vessel interaction accurately enough to influence device selection and reduce complications or reinterventions. Stage 1 therefore centres on verification, validation, workflow fit and the quality-management evidence needed for a future medical-device certification pathway.

19. iNAMed: Computational Toxicology for Medical-Device Regulation

Grant ID 101328775; QSAR Lab Sp. z o.o., Poland. iNAMed is translating AI-driven computational toxicology into medical-device conformity assessment under Regulation (EU) 2017/745. Its QL Materials platform predicts genotoxicity, mutagenicity and inflammatory responses to titanium-dioxide and silicon-dioxide nanoforms.

The industrial case uses a continuous glucose-monitoring sensor to connect computational predictions with a real device and its materials. iNAMed's goal is precedent-setting regulatory use, not another prediction benchmark in isolation. External validation, standards alignment, transparent documentation and a defined integration route for conformity assessment will determine whether the method becomes trusted evidence for notified bodies and manufacturers.

20. Pan-NAM: Human Pancreatic Microtissues for Metabolic Drug Discovery

Grant ID 101327792; STEMX Bio B.V., Netherlands. Pan-NAM benchmarks PancreaScale, a platform producing functional human pancreatic microtissues from induced pluripotent stem cells. The tissues contain beta, alpha and delta cells, reproduce important elements of native islet architecture and respond to glucose while avoiding recombinant growth factors and animal-derived components.

The project targets high-throughput metabolic drug discovery, including response to GLP-1 medicines and modelling of type 2 diabetes. Its Stage 1 challenge is to prove scalable 384-well production, consistent tissue function and meaningful disease or drug-response readouts. The combination of speed, multicellular structure and animal-component-free production is commercially relevant only if performance remains reproducible across plates, batches and users.

What the NAM Portfolio Reveals

The Portfolio Balances Biological and Computational Models

Six projects rely primarily on advanced human biological models: STRIALIS, STEM-SAFE, VANTAGE, SCALE-MYO, OCENTRA and Pan-NAM. Four rely primarily on simulation, digital twins or predictive computation: AI4BirthCare, InSteps, TEVAR-Twin and iNAMed. The division shows that the EIC is treating NAM as an evidence strategy rather than a single technology class.

The Use Cases Cover Decisions, Not Just Experiments

• Drug selection and stratification: STRIALIS seeks to match therapies to mechanistic Parkinson's subtypes.
• Safety testing: STEM-SAFE and iNAMed address gene-therapy and medical-device risks.
• Efficacy and toxicity: VANTAGE evaluates multiple effects of antibody-drug conjugates.
• Drug discovery and screening: SCALE-MYO, OCENTRA and Pan-NAM create scalable human test systems.
• Clinical and device planning: AI4BirthCare, InSteps and TEVAR-Twin model patient or device outcomes.

That spread is important because adoption occurs at different decision points. A NAM used to screen compounds early has different validation requirements from a model intended to support medical-device conformity assessment or a patient-specific clinical decision. The portfolio can test whether the same stage-gated funding logic works across all of them.

Regulatory Readiness Is a Core Technical Output

The main barrier identified by the Challenge is not the absence of scientific methods. It is limited end-user and regulatory confidence. For this reason, regulatory mapping, reproducibility, transferability, standardisation and validation outside the developer's laboratory are part of the work rather than post-project commercial tasks. The strongest Stage 1 evidence will connect assay or model performance to a specific decision a pharmaceutical company, medical-device developer, clinician, regulator or notified body needs to make.

Automation and Scale Recur Across the Wet-Lab Projects

VANTAGE uses an injection-moulded platform and image analytics; SCALE-MYO targets standardised tissue supply; OCENTRA integrates automated culture and screening; Pan-NAM targets 384-well production. These are not incidental engineering features. They address the reproducibility and throughput gap that prevents sophisticated human models from replacing familiar industrial tests.

Country Distribution Across the 20 Selected Projects

The 20 mono-beneficiary projects are coordinated from 10 countries. Germany leads with five projects, all in Physical AI. Italy follows with four, including three NAM projects. Switzerland has three across both Challenges, and the Netherlands has two NAM projects.

The pattern is concentrated but not uniform. Germany and Italy together account for 45% of the portfolio, while six countries have one project each. Switzerland's three selections also show that Horizon Europe Associated Countries are material participants rather than peripheral additions. The cohort includes Central and Southern Europe but no coordinator from a Nordic, Baltic or Balkan country.

Country counts should not be treated as national league-table scores. Selection was performed at proposal and portfolio level, and the EIC has not published complete country-by-country submission counts and evaluation outcomes. The geographic data show where the selected legal entities are established; they do not show where every employee, laboratory, user, partner or future Stage 2 consortium member is located.

What the 20 Projects Must Prove Before Stage 2

The Stage 1 projects have a narrow window from September 2026 through May 2027. Their task is not to finish commercial development. It is to create a defensible basis for selecting the smaller set of solutions that warrant a longer and more expensive real-world development programme.

The Shared Evidence Package

1. A named benchmark. Each project needs a credible comparison against the relevant state of the art. An improvement claim without a baseline, measurement protocol and repeatable result does not satisfy the purpose of Stage 1.
2. Technical viability in the target use case. Laboratory capability must connect to the operational environment described in the proposal.
3. End-user commitment. A letter of intent opened the door, but Stage 2 requires stronger evidence that users, customers, integrators or regulators will contribute to development and testing.
4. An adoption plan. Safety, certification, data governance, standardisation, manufacturability, clinical acceptance and procurement cannot be postponed until after the technology works.
5. A credible Stage 2 work plan. The larger proposal must convert Stage 1 findings into milestones, resources, deployment activities and risk controls.
6. Portfolio value. Projects will again be considered in relation to the other solutions, not solely as isolated grants.

Physical AI Milestones

Physical AI teams must produce initial performance data demonstrating technical viability and disruptive potential against the state of the art in at least one pilot case. They must also show clear commitment from end users and stakeholders to continue development. The anticipated Stage 2 then moves toward robust integrated operation in complex or semi-structured environments, at least two deployment loops, scalability or manufacturability, compliance preparation and TRL 6-7.

For these projects, a polished demo is not equivalent to a benchmark. The evidence needs to cover the metric that makes the solution worth adopting: task success across heterogeneous objects, intervention reliability, handover safety, mission endurance, experimental-cycle reduction, policy transfer or another use-case-specific performance measure.

NAM Milestones

NAM teams must show that their method is viable for human-relevant biomedical innovation and satisfies the needs of the identified use case. They also need clear industrial end-user commitment and a well-defined regulatory plan. Stage 1 activities include small-scale experiments or modelling, performance assessment, benchmarking for human relevance, reproducibility and transferability, and mapping of regulatory, clinical, standardisation and industrial needs.

The anticipated Stage 2 moves toward a functional and scalable NAM prototype ready for regulatory, industrial or clinical uptake. Depending on the method, this can require comparison with animal models or human trials, external-laboratory validation, regulatory-grade data, standardisation and evidence that the method can operate reliably outside the developer's own team.

Stage 2 Consortium Options

Stage 1 was restricted to mono-beneficiaries, but the EIC allows a Stage 2 application from one eligible entity or a small consortium of two or three independent legal entities. New partners can join where they contribute complementary expertise, infrastructure, user access, scale-up capacity or regulatory validation. The Stage 1 legal entity must remain involved or ensure that the new consortium has the rights required to use and commercialise the relevant technology and results.

That flexibility is strategically important. Several Stage 1 projects are led by universities or research organisations and will need a commercial route. Others are led by small companies that may need a hospital, industrial user, integrator, contract research organisation or specialist validation partner. Consortium expansion is useful when it closes a documented execution gap; adding partners simply to make the proposal look larger would make governance and resource allocation harder to defend.

What Applicants, Users and Investors Can Learn From the First Cohort

1. AIC Selected Systems With a Specific Place to Fail or Succeed

The projects do not describe generic platform potential. They name environments and decisions: handling garments in warehouses, cleaning semiconductor burn-in boards, inspecting subsea infrastructure, planning stent-graft placement, assessing gene-therapy tumour risk or screening GLP-1 responses. A strong challenge proposal turns a large strategic theme into a measurable confrontation with reality.

2. TRL 4 Was an Entry Condition, Not the Destination

The programme did not fund ideas that still needed a laboratory proof of concept. Selected applicants already had an initial prototype or validated model. Stage 1 finances comparative validation, risk reduction and adoption planning. Future challenge applicants should therefore build evidence, testing access and user relationships before the call opens.

3. The Demand Side Is Part of Evaluation

Every winning concept had to demonstrate stakeholder interest. During Stage 1, that interest must become operational participation. For industrial users, hospitals, regulators, integrators and procurers, this creates an early opportunity to influence a technology before it is locked into a product design. For projects, it creates a stricter test: a user who supplies data, infrastructure, personnel and acceptance criteria is more valuable than a generic endorsement.

4. Benchmark Design Is a Competitive Asset

The programme repeatedly emphasises comparison with the state of the art. Teams that define the wrong baseline can complete substantial technical work and still fail to demonstrate breakthrough value. Benchmark design should specify the incumbent method, test population or task distribution, metrics, minimum meaningful improvement, failure cases and independent validation conditions before extensive experimentation begins.

5. Regulation and Safety Cannot Be a Final Work Package

Physical AI projects face product safety, AI governance, privacy, cybersecurity and human-interaction risks. NAM projects face method qualification, medical-device or medicinal-product regulation, clinical evidence and data-governance requirements. The selected cohort shows that EIC commercialisation logic now treats these constraints as design inputs. A technology that cannot generate acceptable evidence for its regulator or user is not market-ready regardless of laboratory performance.

6. The First 20 Are a Portfolio, Not 20 Unrelated Grants

Programme Managers are expected to support cross-project learning, user workshops and synergies. The Physical AI projects can share approaches to benchmarking, safety and interoperability. The NAM projects can compare external validation, standardisation and regulatory engagement. This makes collaboration rational even where projects remain competitors for Stage 2: shared methods can improve the quality of the entire evidence base without erasing each project's differentiator.

7. Stage 2 Will Be a Market-Readiness Filter

The initial competition selected credible opportunities. The next competition evaluates what nine months of work changed. Teams need to show advances against competitors, a robust technical foundation, market-creating potential, committed users or customers, investor relevance, execution under uncertainty and a credible plan. The winning narrative in 2027 will therefore be a chain of verified results rather than a restatement of the original ambition.

Frequently Asked Questions

How many EIC Advanced Innovation Challenges projects were selected in 2026?

The EIC selected 20 Stage 1 projects: 10 under Accelerating Physical AI and 10 under Translating Disruptive New Approach Methodologies into Practice.

What was the EIC AIC 2026 success rate?

The overall Stage 1 selection rate was 2.82%, based on 20 selections from 709 proposals. Physical AI had a 2.35% selection rate, with 10 projects from 425 proposals. NAMs had a 3.52% selection rate, with 10 projects from 284 proposals.

How much funding does each selected project receive?

Each project receives a fixed €300,000 Stage 1 lump sum for up to nine months. The complete Stage 1 cohort therefore represents €6 million in EU grant funding.

Are these EIC Accelerator winners?

No. EIC Advanced Innovation Challenges are a separate pilot funding instrument. The projects receive Horizon Europe lump-sum grants and do not receive an EIC Fund equity commitment through this selection.

Does every Stage 1 project receive €2.5 million in Stage 2?

No. Stage 2 is a restricted competition. Stage 1 participation creates eligibility to apply, not an entitlement to further funding. The maximum anticipated Stage 2 grant is €2.5 million for up to 2.5 years.

Can an organisation that did not win Stage 1 apply directly to Stage 2?

It cannot submit an independent Stage 2 proposal. Only projects selected for Stage 1 can enter the restricted call. A new organisation can join a Stage 2 consortium under the applicable eligibility and intellectual-property conditions, but the application must originate from a Stage 1 project and preserve access to the relevant technology and results.

When is the EIC Advanced Innovation Challenges Stage 2 deadline?

The current EIC programme page lists 18 June 2027 at 17:00 Brussels time as the indicative deadline. Applicants must use the adopted 2027 Work Programme and the Funding & Tenders Portal call documents as the controlling sources when the restricted call opens.

What does Physical AI mean in this Challenge?

Physical AI means AI integrated into embodied systems that sense, reason and act in the physical world. The portfolio includes humanoids, autonomous maritime vehicles, exoskeletons, industrial manipulators and self-driving laboratories, as well as enabling data and operating-system infrastructure.

What are New Approach Methodologies?

In this Challenge, NAMs include human organoids, organ-on-chip systems, advanced tissues, digital twins, virtual patients, computational toxicology and AI-enhanced predictive models used in biomedical research or the testing of medicines and medical technologies. Their purpose is to improve human relevance and replace, reduce or refine animal use.

Can EIC AIC projects focus on defence?

No. The official EIC FAQ states that Advanced Innovation Challenges must retain an exclusively civil focus under the general Horizon Europe rules. Results can later have security relevance, but a proposal cannot have a primarily military or defence purpose.

Do unsuccessful Stage 1 applicants receive a Seal of Excellence?

No. The EIC's official Advanced Innovation Challenges FAQ states that no Seal of Excellence is awarded in Stage 1.

Sources and Methodology

This analysis was prepared on 11 September 2026 from official European Commission and CORDIS records. Project names, grant IDs, dates, EU contributions, coordinator countries and objectives were checked against the individual CORDIS fact sheets. Selection-rate, oversubscription and geographic-share calculations are derived from the official totals and the 20 published project records; percentages are rounded to two decimal places where appropriate.

• EIC announcement: first Advanced Innovation Challenges projects
• EIC submission statistics: 709 proposals from 39 countries
• EIC Advanced Innovation Challenges pilot page
• Official EIC Advanced Innovation Challenges FAQ
• Adopted EIC Work Programme 2026
• CORDIS portfolio: HORIZON-EIC-2026-AIC-01 Physical AI
• CORDIS portfolio: HORIZON-EIC-2026-AIC-02 New Approach Methodologies

Accuracy note: Stage 2 provisions in the adopted 2026 Work Programme are indicative and subject to the adopted 2027 Work Programme. The official call documents in force at submission are authoritative.