Building the shared training ecosystem for better weeding robots.
We bring field images, verified annotations, AI models and robot integrations into one farmer-controlled W3DS environment. The goal is to make robotic weed recognition more accurate, interoperable and affordable across real field conditions.
Weeding robots can already see the field. The problem is that they are still learning alone.
Field images and annotations are scattered across farms, robots, research projects and proprietary platforms. The same crops and weeds are labelled repeatedly, while valuable datasets remain difficult to find, compare and reuse.
A model trained on one farm may lose accuracy when the crop, weed species, soil, growth stage, camera or geography changes. Robot manufacturers must divide their effort between hardware and AI, while farmers are asked to invest in machines that are not yet reliable enough across the conditions they face.
AI Weed Library turns these isolated efforts into one collaborative learning cycle.
One ecosystem. Four reusable assets.
Every participant contributes what they do best.
Farmers and robots capture the field. Farmers, students and agronomists annotate and validate plants. AI developers train and compare recognition models. Robot manufacturers select and deploy the models that fit their machines and operating conditions. W3DS records permissions, provenance, quality and reputation throughout the chain.
The result is not one closed platform. It is an interoperable market for data, expertise, models and robot services.
Project targets
Meet PLANT-AI
PLANT-AI — Plant Library & Annotation for Next-generation Training of AI — is the user-facing software suite developed within the project.
It will help users find agricultural images, inspect provenance and permissions, annotate crops and weeds, validate contributions, prepare datasets for training, compare models by context and connect selected models to compatible robot systems.
Built by farmers, data specialists and robotics participants
The formal project consortium brings together Stichting Post-Platforms Foundation, Difco International B.V., J.C.J. Ruiter-Wever, Maatschap Broersen and PuurBio B.V. A wider implementation ecosystem contributes software, AI expertise, field operations and robot integrations. Each role is presented transparently on the Partners page.
Your contribution may be the missing part of a better model.
Do you have field images, existing annotations, crop or weed expertise, an AI recognition pipeline, a robot platform or a farm for validation? Join the project and help build a reusable training ecosystem that works across organisations and machines.
A practical route from fragmented field data to reliable weed-recognition AI
AI Weed Library is the public-facing name of W3DS for AI Weed Detection — W3DS voor AI onkruidbestrijding — an EIP Noord-Holland 2025 project led by Stichting Post-Platforms Foundation.
The project runs from 31 March 2026 to 30 December 2027 and develops a shared, decentralised ecosystem for images, annotations, AI models and robot integration.
Weed control is becoming a structural challenge.
The number of available herbicides is decreasing, while organic growers cannot use chemical herbicides at all. Manual weeding is expensive and increasingly difficult to organise at scale. Lightweight autonomous robots offer a promising alternative, but their field performance depends on the quality and generalisability of their recognition models.
The current development process is inefficient. Farms and robot companies repeatedly collect and label similar images. Datasets remain fragmented. Models are hard to compare. A model that performs well in one crop or soil may fail in another.
The project addresses this bottleneck by making data, annotations and models reusable across organisations while preserving ownership, permissions and provenance.
From isolated projects to a shared learning cycle
Instead of asking every robot manufacturer to build the entire data and AI stack alone, the project separates the value chain into specialist roles. Farmers and robot operators provide real field data. Annotators contribute agricultural knowledge and scalable labelling. AI specialists focus on training and evaluation. Robot manufacturers focus on machines, integration and field performance. W3DS connects these contributions without centralising all data in one platform. This allows the ecosystem to improve collectively while every participant remains identifiable, accountable and able to control the use of their contribution.
The project has four primary objectives
Nine work packages
From one field image to a better model on many robots
AI Weed Library organises a transparent contribution chain. Every digital object remains linked to its owner, context, permissions and quality history — from capture and annotation to training, benchmarking and deployment.
Eight-step chain
Every role, a different vantage point
There is no universally best image, annotation or model.
A useful choice depends on the task. A model for young onions on sandy soil in North Holland may not be the best model for flower bulbs, clay soil or a later growth stage. AI Weed Library therefore makes context a first-class part of the data.
A contribution should remain visible after it enters an AI pipeline.
W3DS links images, annotations, training runs, models and deployments. This makes it possible to see which contributions were used and where value was created.
The project develops the technical and governance foundation for fair compensation. It does not assume that all data must be free, and it does not lock contributors into one marketplace or payment provider.
Data should not belong to applications.
Web 3.0 Data Space separates data from platforms. Images, annotations and models stay in eVaults controlled by their owners, while different applications and services compete to work with that data under permission.
Four components
Pilot-ready eVaults and the W3 Adapters for robots and external systems are active work inside this project — see Work Package 2 and Work Package 3 on the Project page — not a finished catalogue of integrations.
Decentralized architecture
Not one large database — a federated view over authorised, owner-controlled eVaults.
What W3DS prevents
Different systems do not need one identical internal database.
A writing application describes what it stores in a machine-readable form. A reading application uses that description, shared vocabularies and a W3 Adapter to reconstruct the data for its own workflow. This makes interoperability a continuous capability rather than a one-time agreement on one rigid format. In the AI Weed Library, the same image and annotation can therefore support different annotation tools, AI pipelines and robot platforms.
Access is permissioned, not universal.
The fact that many applications can technically connect to W3DS does not mean that every application can access every object. Access control, identity, signatures and auditable provenance are part of the architecture.
For example: an AI developer's training environment can read the images and annotations a farmer has explicitly authorised for that use case — not every image in every other contributor's eVault. A robot manufacturer's W3 Adapter can write a deployment record to the eVaults it has been granted access to, and no others.
W3DS has nothing to do with blockchain.
The architecture uses decentralised ownership, linked data, identity, access control and interoperable services without requiring a blockchain or token economy.
PLANT-AI
Plant Library & Annotation for Next-generation Training of AI
The user-facing workspace for discovering field images, creating and validating annotations, assembling training datasets, comparing AI models and connecting them to agricultural robots through W3DS.
PLANT-AI is being developed within the active EIP Noord-Holland 2025 project. The website must distinguish clearly between planned functionality, pilot releases and publicly available modules. The underlying W3DS-Agro initiative is available at w3ds-agro.eu.
Reference: field map and ownership view from the underlying W3DS-Agro platform, not a confirmed PLANT-AI screenshot.
Functional modules
Four connected data layers
Provenance graph
Explains the future basis for recognition and compensation — without promising an automatic payment.
Roadmap 2026–2027
A project is credible when its targets, evidence and limitations are visible.
This page tracks what the project has committed to deliver, what is currently in progress and what has been completed or validated.
Target dashboard
Coverage matrix — crops × weeds (illustrative)
Shows gaps rather than one total only. Populated from CMS once field collection begins.
Deliverables by work package
Progress presentation rules
- Each deliverable has owner, planned date, status, last update, public summary and a link to the available artefact.
- Counters update from the CMS, not hard-coded into the design.
- Quality indicators always show metric definition, context, sample size and test dataset.
- Recognition accuracy and mechanical removal accuracy are never merged into one figure without a formula.
- Negative or inconclusive results are published as learning, where grant and IP constraints allow.
A shared AI ecosystem needs complementary roles.
The project combines farming practice, project coordination, W3DS architecture, software development, annotation, AI and robot integration. Formal grant beneficiaries and wider implementation participants are shown separately.
Formal consortium members
Wider project ecosystem
Use one of three labels on every partner card: Formal consortium member, Contracted implementation partner, or Wider project ecosystem. The generic label "project partner" is never used for all organisations alike.
Open invitation to new contributors
If your organisation can contribute field data, annotation capacity, AI expertise, robot integration or dissemination reach, we want to hear from you.
Project updates from the field, the data library and the development team
Follow new datasets, annotation campaigns, robot integrations, model benchmarks, demonstrations, publications and lessons learned throughout the project.
Help build the training ecosystem that no single farm or robot company can build alone.
We welcome contributors with field images, annotations, agricultural expertise, AI models, training infrastructure, robot platforms, pilot fields and dissemination networks.
Contribution pathways
What contributors can expect
Thank you for contacting AI Weed Library.
We will review your proposal in relation to the project scope, data rights, technical compatibility and current work packages. Submission does not automatically create a partnership or grant access to project data.
Send your contribution proposal
Contact the project
Hilvertsweg 275
1214 JG Hilversum
The Netherlands
For contributions, please use the Join form so that your proposal reaches the correct work package.