rFabric

Documentation

Architecture Overview

rFabric is organized as a 6-plane backbone plus 10 product components. The backbone provides shared state, control, governance, and economics. The components own the robotics workflows teams actually operate.

Backbone Planes

The backbone is shared across every lifecycle component. It is what makes rFabric one connected robotics system rather than a stack of point solutions with fragile handoffs.

Product Components

The lifecycle groupings below are navigational only. The real architectural unit is the component that owns a durable object set or decision surface.

01 - Data Foundation

Become the system of record for robot data, labels, quality decisions, and training-ready dataset versions.

  • Data Ingestion & Preprocessing
  • Annotation & Labeling
  • Dataset Curation & Versioning

02 - Model Development

Turn curated datasets into reproducible runs, evaluated candidates, and promoted model state with explicit evidence.

  • Model Training & Evaluation
  • Model Registry

03 - Release Management

Package approved models into deployable artifacts and roll them out through governed, reversible operational workflows.

  • Release Packaging
  • Deployment & Updates

04 - Live Operations

Operate fleets safely in the field and turn deployment outcomes into new data, new evaluation cases, and better future models.

  • Fleet Management
  • Monitoring, Alerting & Maintenance
  • Human Intervention & Robot Control

Cross-Cutting Contracts

The architecture matters because every component writes into the same entity graph and uses the same control model. That keeps handoffs explicit even as teams adopt more of the platform over time.

Shared entity ownership

Components do not create private metadata islands. Datasets, training runs, models, artifacts, deployments, telemetry, and interventions all exist in one canonical graph.

Control plane and data plane separation

Heavy binary movement happens on the data plane, but state transitions, lineage, approvals, and policy remain authoritative in the control plane.

Promotion-boundary immutability

Curated datasets, promoted models, packaged artifacts, and finalized operational or financial records are versioned instead of mutated in place.

Operations feed development

Intervention sessions, field anomalies, and maintenance outcomes are new data and evaluation input, not terminal dashboards.

Why This Architecture Fits Robotics

Robotics data is multimodal and time-aligned

The platform has to model synchronized sensor streams, derived assets, labels, and quality decisions directly instead of pretending everything is a log line or a file blob.

The deployable artifact is probabilistic behavior

Robotics versioning has to connect datasets, curation ruleset, training context, release packaging, and field outcomes rather than stopping at code commits.

Deployment is physical, not abstract

Rollout, rollback, OTA delivery, fleet health, maintenance, and teleoperation are operating workflows with safety and logistics implications, not just CI/CD steps.

Human-in-the-loop is structural

Review, approval, escalation, and intervention are core parts of robotics workflows. The architecture keeps them first-class instead of treating them as side-channel exceptions.

Design Rules

Robotics-native data model

Robots produce multimodal, time-aligned sensor streams and probabilistic artifacts. The platform has to model that reality directly.

Human gates are first-class

Review, approval, teleoperation, and escalation belong inside the platform rather than in chat threads and side channels.

Structural isolation

Workspace ownership, organization policy, regions, and environments are architecture primitives, not admin afterthoughts.

Field-to-lab compounding loop

Deployment outcomes, telemetry anomalies, and intervention sessions are development inputs, not operational dead ends.