Skip to content

Synthex for Autonomous Systems

Prepare Autonomous AI for the Unexpected

Explore configurable scenarios, synthetic perception datasets and model evaluation workflows designed around difficult physical-world conditions.

Built with these teams in mind

  • Autonomous vehicle teams
  • Robotics companies
  • Computer vision engineers
  • Simulation engineers

Simulation supports development and testing; it does not establish real-world safety or certification.

Industry challenges

The long tail is where perception fails

Perception systems meet situations that are rare, dangerous or expensive to capture. Controlled scenarios make them repeatable.

  • Rare and difficult events

    Critical situations may appear a handful of times in thousands of hours of driving or operation.

  • Adverse weather

    Rain, fog and spray change appearance and sensor returns in ways clear-weather data does not cover.

  • Poor visibility

    Night, glare and low contrast reduce what cameras can resolve.

  • Unusual object positions

    Objects partly hidden, fallen over or in unexpected places break learned assumptions.

  • Sensor variation

    Different mounting, calibration and noise change what a model receives.

  • Limited scenario coverage

    Teams often cannot say which conditions their test data actually covers.

  • Expensive real-world data collection

    Fleets, staging and labelling make each new scenario costly to capture.

Core capabilities

Scenarios, data and evaluation for perception teams

Planned capabilities. They connect to the Simulation, Data and Evaluation engines.

  • Platform Vision

    Scenario Design

    Describe actors, conditions and success criteria as structured, versioned scenarios.

    Example output
    A defined scenario collection for a perception task.
  • Platform Vision

    Virtual Environment Configuration

    Assemble roads, junctions and yards from configurable layouts and assets.

    Example output
    Reusable environment configurations.
  • Platform Vision

    Synthetic Perception Data

    Capture images and labels across controlled variations of each scenario.

    Example output
    A synthetic perception dataset sample.
  • Platform Vision

    Object and Sensor Annotation

    Boxes, masks and sensor metadata derived from known scene ground truth.

    Example output
    Labels with per-frame sensor metadata.
  • Platform Vision

    Edge-Case Evaluation

    Test models on hard, rare scenarios and group failures by cause.

    Example output
    An edge-case analysis with failure categories.
  • Platform Vision

    Repeatable Benchmarking

    Re-run identical scenario suites as models change and compare like for like.

    Example output
    Versioned benchmark results over time.

How to read status labels

Platform Vision
A planned capability that defines where the platform is heading. Not yet available.
Research Direction
Active exploration of methods and questions. No prototype or results yet; not a product.
Illustrative Demonstration
A visual or interactive explanation using sample content. No live system or real results behind it.
Validated Implementation
Implemented and verified against defined criteria. No capability on this site is presented at this level yet.

Interactive explorer

Autonomous Perception Scenario Explorer

Combine an environment, weather, lighting and detection task to see how a perception scenario could be described. Educational illustration: no simulator is executed.

Illustrative Interface — No Simulation Executed

Illustrative Demonstration
Environment
Weather
Lighting
Scenario
ILLUSTRATION · NOT A SIMULATION OUTPUT

intersection.rain.dusk.pedestrian

Intersection · Rain · Dusk · Pedestrian Detection

Scenario summary: Pedestrian Detection in an intersection scene, rain weather, dusk lighting.

Expected annotation types

  • 2D bounding boxes
  • Instance masks
  • Pose or keypoints (optional)
  • Per-frame scenario metadata

Example evaluation concerns

  • Small, partly hidden or unusually posed pedestrians
  • Many agents crossing paths at different distances
  • Streaks, spray and reflections on wet surfaces
  • Low sun glare and long shadows

Educational illustration. It does not run CARLA, Unreal Engine, Unity or any physics simulation, and it produces no data. Simulation-based testing does not by itself establish real-world safety.

Simulation workflow

From scenario definition to failure analysis — and back

The intended loop for scenario-based perception testing. Failure analysis feeds the next round of scenarios.

Capture is conceptual

No simulation backend is integrated yet. Steps marked “Conceptual” describe intended behaviour.
  1. 01

    Scenario Definition

    State the situation, actors and what the perception system must detect.

  2. 02

    Virtual Environment Setup

    Load or assemble the environment layout.

  3. 03

    Object and Condition Variation

    Vary agents, positions, weather and lighting within defined ranges.

  4. 04

    Virtual Sensor CaptureConceptual

    Record camera and other sensor outputs with exact ground truth.

  5. 05

    Synthetic Dataset Preparation

    Validate labels and package the captured data.

  6. 06

    Perception Model Evaluation

    Evaluate per scenario, condition and object type.

  7. 07

    Failure Analysis

    Group errors by cause and inspect representative frames.

Scenario Refinement · returns to step 01

Failure categories define new, harder scenario variations for the next cycle.

Sensor data concepts

Data representations for perception

Representations a scenario-based workflow would work with. None is presented as actively supported: no capture pipeline is implemented yet.

  • Platform Vision

    RGB Cameras

    Colour images from configurable virtual cameras: placement, field of view, exposure.

  • Platform Vision

    Depth Data

    Per-pixel distance from the camera, known exactly in simulation.

  • Platform Vision

    LiDAR

    Point clouds from modelled laser scanners; realism depends heavily on the sensor model.

  • Platform Vision

    Segmentation Masks

    Pixel-level class and instance labels aligned to each camera frame.

  • Platform Vision

    Bounding Boxes

    2D and 3D object boxes with class, occlusion and truncation attributes.

  • Platform Vision

    Scenario Metadata

    Weather, lighting, actor positions and sensor settings recorded per frame.

Edge-case examples

Situations worth testing deliberately

Illustrative edge cases and what a perception team could evaluate in each. Drawings are schematic, not simulator output.

  • Edge case 01

    Pedestrian partially occluded

    What to evaluate

    Detection rate and box accuracy as visible area shrinks behind parked vehicles.

  • Edge case 02

    Vehicle in heavy fog

    What to evaluate

    Detection range and confidence as contrast falls with distance.

  • Edge case 03

    Night-time obstacle detection

    What to evaluate

    Missed low obstacles outside headlight cones; false alarms from reflections.

  • Edge case 04

    Complex intersection

    What to evaluate

    Tracking many agents crossing paths; confusion between overlapping objects.

  • Edge case 05

    Unusual roadside object

    What to evaluate

    Behaviour on classes absent from training, such as fallen loads or debris.

  • Edge case 06

    Low-visibility industrial environment

    What to evaluate

    Detection of people and forklifts in dust, steam or poor lighting.

Technology

Possible simulation technologies

Synthex is evaluating established simulation tools as possible integrations or research tools behind a common scenario layer. No integration is complete, and there is no partnership with or endorsement by their owners.

Synthex scenario layer (planned)

Backend-independent scenario definitions, variation sweeps, capture settings and hand-off to the Data and Evaluation engines.

Platform Vision
  • CARLA

    Under exploration

    Open-source simulator for autonomous driving research with configurable maps, traffic and sensors.

  • Unreal Engine

    Under exploration

    Real-time 3D engine from Epic Games, suited to photorealistic environments.

  • Unity

    Under exploration

    Real-time 3D platform with tooling usable for synthetic data generation.

  • Other frameworks

    Under exploration

    Robotics-oriented simulators such as Gazebo may suit indoor and industrial scenarios.

CARLA, Unreal Engine, Unity and Gazebo are names or trademarks of their respective owners and are mentioned only to describe technologies under consideration.

Pilot positioning

What a focused autonomous systems pilot could look like

Example answers to the questions a perception or simulation lead would ask. Deliverables are examples of what a scoped pilot could produce — not an operational service.

01What problem does the team face?
Perception models that have not been tested on the rare, adverse or unusual situations they will meet.
02Why is it difficult to solve?
Edge cases are rare, dangerous or expensive to stage, and real test data has unclear scenario coverage.
03How could Synthex address it?
Defined scenario collections, controlled variation, synthetic perception data and edge-case evaluation.
04What could a focused pilot produce?
Example deliverableA defined scenario collection, a synthetic perception dataset sample and an edge-case analysis.
05What evidence would show technical value?
Repeatable per-scenario results, failure categories traced to conditions, and checks against real held-out data.

Autonomous systems pilot

Explore Safer, Repeatable Autonomous Perception Testing

We are interested in working with perception and simulation teams on scenario design, synthetic data and edge-case evaluation.

Simulation-based testing can help find weaknesses, but it does not by itself establish real-world safety or meet any certification requirement.