Unified data infrastructure
GitHub Let's talk

Autonomous Systems

Vehicles, sensors and decisions, close to the edge.

An autonomous system produces a decision trail: what was sensed, what was inferred, what was chosen and what a human did about it.

See Real-time Analytics All of Infrastructure & Technology

An incident review that is a query over the stream, the inference and the intervention.

The story

An incident review that is a query over the stream, the inference and the intervention.

Where it starts

Sensor streams

Lidar, radar, camera and inertial measurements. It is the first of 4 workloads running in Autonomous Systems.

The question it raises

Incident reconstruction

The stream, the inference record and the intervention stay in one layer, so a review is a query.

Why one question is hard

From Sense to Review

Autonomous Systems data moves through 4 stages — Sense → Infer → Act → Review. The shapes in play are Time series, JSON / documents, Objects, SQL, Events, and answering one question means reading across all of them.

What PLOMID contributes

Autonomy produces evidence at machine speed. These are the parts that keep it reviewable.

The environment

Sensor streams, perception output, decisions, interventions and logs.

Sensor streams are high-rate and time-ordered, perception output is structured per frame, and interventions and logs are events that must be reconstructable. PLOMID keeps the stream, the inference record and the intervention in one layer, so an incident review reads from one source.

Workload architecture

The system reading itself.

Planning, execution and transactions first — then the workloads that use them and the shapes they name.

Autonomous Systems · workload architecture
Workloads

What runs against this data.

  • Sensor streams
  • Perception and inference output
  • Decisions and interventions
  • Fleet and version records
Data models

The shapes those workloads read and write.

  • Time series
  • JSON / documents
  • Objects
  • SQL
  • Events
The layer

One path from a request to the data it names.

  • Planning Predicates narrow the work before it runs
  • Execution Records, fields and windows answered together
  • Transactions Readers and writers do not block each other
Surfaces

How the work reaches the layer.

  • SQL surface The query language the layer is documented in
  • Applications Services and jobs writing and reading as they run
  • Analytics & AI clients The same layer, the same access path
One environment · many workloads

What runs against autonomous systems data.

4 workload families over one set of shapes. Choose one to see what it moves and where it lands.

Lidar, radar, camera and inertial measurements.

  • Planned once against the layer, not once per store
  • Read beside the records it shares a key with
  • Persisted under one storage contract
The data journey

How autonomous systems data reaches one layer.

Walk the path the data takes, from the environment that produces it to the questions it answers. Select a station, or a shape, to read each step.

An incident review that is a query over the stream, the inference and the intervention.

Environment

The vehicle

A unit in the field with sensors, software versions and calibration.

SQL

Data models in play

The shapes, in one layer.

5 shapes carry this domain. Choose a stage to read the operation, or a shape to see every stage that handles it.

PLOMID · Autonomous Systems sense · infer · act · review Select a stage
Stage

Sense

Sensor streams and synchronised frames

Time series · Objects

Workload map Autonomous Systems workload map. Every shape on it is a surface of the layer, and each stage names the part of the operation it carries.
  • Time series Measurements and events in time order
  • JSON / documents Documents and nested objects
  • Objects Large assets with queryable metadata
  • SQL Records, keys and joins
  • Events Operational events as they happen
Where the data goes to work

Questions a fleet asks about itself.

Each one reads telemetry with the records that explain it, from the same layer rather than a sidecar store.

Incident reconstruction

The stream, the inference record and the intervention stay in one layer, so a review is a query.

  • Time series
  • Objects

Version attribution

Model and software versions are recorded per run, which is what makes behaviour changes attributable.

  • JSON / documents
  • Vector

Fleet-level metrics

Disengagements and interventions compute over the events that produced them.

  • Events
  • SQL

Offline then reconcile

Data recorded without a link is reconciled into the same layer rather than a separate store.

  • Objects
  • SQL
Deployment & residency

Where this data is allowed to run.

Vehicles and cells operate without a reliable link, so local recording and later reconciliation are the norm.

Deployment, residency and control
What you build next

Real-time Analytics

Windows and aggregates over data that is still being written.

If Incident reconstruction is your question, start here.

See Real-time Analytics Every solution
Next

Where to go next

Solutions The workload answers this industry leans on, with their models and status. Platform The data models in the product today, and the ones extending the layer. Architecture How a request is planned, executed and persisted once it reaches the layer.