Skip to content

Synthex Research & Innovation

Advancing the Foundations of Physical AI

Exploring synthetic data generation, virtual-world simulation, annotation quality and reproducible evaluation for real-world AI applications.

Research directions and methods. No Synthex experimental results are published yet; every claim is traceable in our evidence register.

Research mapConcept
Concept map: medical imaging, agriculture and autonomous systems research each draw on Synthex data, simulation and evaluation capabilities.

Research themes

Four questions underpin the platform

Each theme pairs a scientific question with an engineering one. Answers are not assumed; they are what the research is for.

  • THEME 01

    Synthetic Data Generation

    When does generated data add information a model can use, and when does it add bias or noise?

    Open questions

    • Which variations matter for a task?
    • How large is the synthetic-to-real gap?
    • When does synthetic data stop helping?
  • THEME 02

    Simulation and Scenario Design

    How should scenarios be specified so they are representative, repeatable and comparable?

    Open questions

    • How to describe coverage formally?
    • Which parameters drive failures?
    • How faithful must sensors be?
  • THEME 03

    Annotation Quality and Dataset Engineering

    How can label errors be found and measured before they distort training and evaluation?

    Open questions

    • Which signals predict bad masks?
    • How should review be prioritised?
    • How should disagreement be recorded?
  • THEME 04

    Model Evaluation and Robustness

    How can evaluation reveal where and why models fail, not just how often?

    Open questions

    • Which metrics suit which tasks?
    • How to measure degradation fairly?
    • How to keep results reproducible?

Featured research projects

Three domains, one research method

Each project shows its objective, approach and actual evidence status. None is labelled complete, because no completed results are documented yet.

  • Medical imaging

    Evidence: Conceptual · R-006

    Synthetic augmentation and segmentation evaluation

    Research objective
    Understand whether controlled synthetic variation can support lesion segmentation research — and when it does not.
    Problem statement
    Annotated lesions are scarce and inconsistent, and models can degrade under acquisition and scanner variation.
    Technical approach
    Fixed patient-level splits, real-data baselines, augmentation as an addition, and evaluation on held-out real images only.
    Relevant engines
    Data EngineEvaluation Engine
  • Agriculture

    Evidence: Conceptual · R-007

    Synthetic crop and weed datasets

    Research objective
    Investigate how generated field imagery with pixel-level labels could extend scarce agricultural datasets.
    Problem statement
    Crops and weeds look alike, field conditions vary widely, and pixel labels are slow to produce.
    Technical approach
    Parameterised field scenarios, mask validation, and evaluation grouped by lighting, growth stage and field.
    Relevant engines
    Simulation EngineData EngineEvaluation Engine
  • Autonomous systems

    Evidence: Conceptual · R-009

    Simulation-based perception testing

    Research objective
    Study how controlled virtual scenarios can probe perception models under rare and adverse conditions.
    Problem statement
    Edge cases are rare, unsafe or costly to capture, and real-world tests are hard to repeat exactly.
    Technical approach
    Scenario specifications, controlled variation, synthetic sensor capture and failure analysis with real-data validation.
    Relevant engines
    Simulation EngineData EngineEvaluation Engine

Research lifecycle

From question to reproducibility review

Every project follows the same lifecycle, so results can be compared, repeated and challenged.

  1. 01

    Research Question

    A specific, falsifiable question with a defined success criterion.

  2. 02

    Dataset and Scenario Selection

    Licensed data and scenario definitions, with provenance recorded.

  3. 03

    Experiment Design

    Baselines, controls, splits and metrics fixed before results are seen.

  4. 04

    Data Generation or Processing

    Versioned generation or preprocessing with recorded parameters.

  5. 05

    Annotation Validation

    Automated checks and expert review of labels.

  6. 06

    Model Training or Evaluation

    Fixed configurations and documented random seeds.

  7. 07

    Results Analysis

    Per-condition results, error analysis and negative findings.

  8. 08

    Reproducibility Review

    Re-run from the manifest; confirm results before any claim.

Decided before results

Baselines, splits, metrics and success criteria are fixed at the experiment-design stage, so results cannot quietly reshape the question.
Read the full methodology

Publications and outputs

Research artefacts we intend to publish

This section lists only real, approved outputs. None have been published yet, so it describes the artefacts each project is designed to produce.

No research outputs have been published yet.

  • Intended artefact

    Technical reports

    Methods, results and limitations for each completed study, including negative findings.

  • Intended artefact

    Research publications

    Peer-reviewed or preprint papers where work is mature enough to be scrutinised.

  • Intended artefact

    Open datasets

    Licensed synthetic or derived datasets with documentation and dataset cards.

  • Intended artefact

    Reproducible notebooks

    Code and configurations that re-create a published result from its manifest.

  • Intended artefact

    Benchmark results

    Versioned scenario suites with fixed protocols and comparable results.

  • Intended artefact

    Project documentation

    Experiment manifests, metric definitions and data governance notes.

Evidence and provenance

How we decide what to say

Every research statement on this site is recorded in an evidence register with its source, verification class and publication decision.

Evidence: Verified
5 entries · Supported by a reviewable source.
Evidence: Experimental
0 entries · Work exists, but claims or results need further verification.
Evidence: Conceptual
4 entries · A proposed workflow, architecture or research direction.
Evidence: Unavailable
11 entries · No source evidence currently exists.
Evidence register: claims, verification class and public presentation decision
IDClaim
R-001Synthex is developing synthetic data, simulation, annotation and evaluation infrastructure for physical-world AI.
Evidence: Conceptual

Published as a statement of intent and direction, never as an available product.

R-002LIDC-IDRI is a public thoracic CT collection with radiologist lesion annotations, hosted by The Cancer Imaging Archive.
Evidence: Verified

Cited as public research context with attribution. No images from the collection are reproduced.

Public source: cancerimagingarchive.net
R-003Synthex lung CT lesion segmentation results (Dice, IoU, precision, recall or sensitivity values).
Evidence: Unavailable

Not published. Pages show the comparison protocol with 'no verified result' cells.

R-004Prior academic or dissertation work on lung CT segmentation and synthetic augmentation.
Evidence: Unavailable

Not described as Synthex research and not referenced on public pages.

R-005Synthetic augmentation improves lesion segmentation performance.
Evidence: Unavailable

Presented only as an open research question, including the possibility that it does not help.

R-006Medical imaging research workflow: preprocessing, masks, augmentation, training, metric evaluation, error analysis.
Evidence: Conceptual

Published as proposed research methodology, not clinical deployment.

R-007Synthetic crop and weed dataset research for agricultural perception.
Evidence: Conceptual

Published as a research direction with schematic illustrations.

R-008Agricultural segmentation benchmark values (IoU, Dice, precision, recall by condition).
Evidence: Unavailable

Not published.

R-009Simulation-driven autonomous perception research (scenarios, synthetic sensor data, robustness evaluation).
Evidence: Conceptual

Published as a research direction.

R-010Integration with CARLA, Unreal Engine, Unity, AirSim or another simulator.
Evidence: Unavailable

Simulators described only as technologies under exploration; no integration or partnership claimed.

R-011CARLA is an open-source simulator for autonomous driving research, built on a fork of Unreal Engine; code MIT, assets CC-BY.
Evidence: Verified

Described factually as a candidate research tool; no affiliation implied.

Public source: carla.org
R-012Descriptions of Unity, AirSim and Gazebo used on Sprint 2–3 pages.
Evidence: Unavailable

Retained as generic descriptions; flagged for owner verification. Research pages name only CARLA and Unreal Engine.

R-013Synthex publications, technical reports or preprints.
Evidence: Unavailable

No entries displayed; output types described as intended artefacts. No scholarly structured data.

R-014Open datasets, reproducible notebooks or public benchmark results from Synthex.
Evidence: Unavailable

Not displayed.

R-015Interactive research demonstrations use static, original fixtures (schematic drawings and descriptive text).
Evidence: Verified

Published with 'Illustrative · static fixture' labels. No backend, generation or inference.

R-016Definitions of Dice, IoU, precision, recall and sensitivity.
Evidence: Verified

Published as definitions only, with the caveat that technical metrics do not establish clinical effectiveness.

R-017Research partners, collaborators, customers or institutional affiliations.
Evidence: Unavailable

None claimed. Collaboration is offered, not asserted.

R-018Regulatory approval, clinical validation, medical device certification or safety certification.
Evidence: Unavailable

Explicitly disclaimed on medical and autonomy pages.

R-019Data Engine copy stating the direction was 'informed by work with public research datasets such as LIDC-IDRI'.
Evidence: Unavailable

Reworded in Sprint 4: the wording implied prior work for which no evidence exists in the repository.

R-020No private research assets, patient data, credentials or proprietary datasets are published by the website.
Evidence: Verified

Guard retained in the e2e suite.

Collaborate

Collaborate on the Next Generation of AI Data Infrastructure

We are looking for collaborators who care about rigorous, reproducible evidence as much as we do.

  • Academic research collaborationJoint questions on synthetic data, simulation and evaluation methods.
  • Technical evaluation projectsStructured evaluation of an existing model or dataset.
  • Dataset research partnershipsLicensed datasets, annotation schemas and scenario libraries.
  • Industry proof-of-concept discussionsScoped, time-boxed exploration of a specific technical problem.