Autonomous Data Loop

简体中文

An autonomous driving data loop system design and engineering portfolio, covering vehicle-cloud data upload, MCAP playback, automated annotation, 3D quality inspection, dataset export, training, evaluation, and feedback-driven data collection.

This repository is a sanitized technical portfolio. It focuses on architecture, engineering decisions, staged delivery, demo artifacts, and reusable examples. It does not include proprietary source code, internal endpoints, credentials, customer data, or confidential deployment details.

What This Project Shows

This project demonstrates how an autonomous driving data loop can evolve from basic data upload to a complete engineering loop:

1. Collect and package vehicle-side data.
2. Upload selected data from edge or field collection nodes to the company-side platform.
3. Review and replay MCAP assets on the platform.
4. Run automated annotation and 3D quality inspection.
5. Export labels in multiple dataset formats.
6. Build training datasets and model versions.
7. Run automated evaluation and feed failed cases back into the next data collection cycle.

System Roadmap

Phase	Name	Scope	Status
Phase 1	Vehicle-cloud data upload	Metadata, probe rules, MCAP packaging, compression, upload, cloud ingestion, playback verification	Completed
Phase 2	Automated annotation and QC	MCAP asset management, automated annotation, Xtreme1 3D quality inspection, label versioning, KITTI/nuScenes export	Designed, ready for implementation
Phase 3	Dataset management	Canonical labels, dataset versioning, sample selection, dataset traceability	Planned
Phase 4	Model training	Training task orchestration, model registry, metric tracking, model-to-data traceability	Planned
Phase 5	Automated test and evaluation	MCAP replay, SOC/Xavier execution, perception output comparison, evaluation reports	Planned
Phase 6	Feedback-driven collection	Failed-case analysis, probe-rule updates, targeted recollection, closed-loop iteration	Planned

High-Level Architecture

flowchart LR
  A["Test Vehicles"] --> B["Vehicle Data Capture"]
  B --> C["MCAP Recording and Probe Rules"]
  C --> D["Edge Node Packaging"]
  D --> E["Public Transfer / Shared Storage"]
  E --> F["Company Platform"]
  F --> G["MCAP Replay and Selection"]
  G --> H["Automated Annotation"]
  H --> I["Xtreme1 3D Quality Inspection"]
  I --> J["Canonical Labels"]
  J --> K["KITTI / nuScenes Export"]
  J --> L["Training Dataset"]
  L --> M["Model Training"]
  M --> N["Automated Evaluation"]
  N --> O["Failed Cases and Probe Updates"]
  O --> C

Loading

Repository Structure

autonomous-data-loop/
├─ docs/                 # Bilingual technical documentation
├─ diagrams/             # Architecture diagrams and flow diagrams
├─ demos/                # Videos, screenshots, and interactive demo pages
├─ examples/             # Sanitized metadata, manifest, API, and label examples
├─ src-demo/             # Small public demo code, not production code
├─ assets/               # Shared images and media assets
├─ README.md             # English entry point
└─ README.zh-CN.md       # Chinese entry point

Documentation

All documentation is maintained in both English and Chinese:

• English files use README.md or *.md.
• Chinese files use README.zh-CN.md or *.zh-CN.md.

Start here:

• Overview
• Phase 1: Vehicle-cloud Data Upload
• Phase 2: Automated Annotation and QC
• Architecture
• AI Agent Workflow
• Phase 3: Dataset Management
• Phase 4: Model Training
• Phase 5: Automated Test and Evaluation
• Roadmap

Public Code Strategy

The real production code is not published here. Instead, this repository will include sanitized and reusable demo code:

• MCAP metadata indexing examples.
• Probe rule examples.
• Package manifest examples.
• Label schema examples.
• Canonical label to KITTI / nuScenes conversion demos.
• Task state machine demos.
• Mock API definitions for platform workflows.

This keeps the repository useful for technical review while protecting company assets and deployment details.

AI-Assisted Engineering

This project also records how AI agents can assist real engineering work:

• Turning raw project materials into structured design documents.
• Iterating architecture after review feedback.
• Generating presentations, demos, and technical documentation.
• Breaking down implementation plans and schedule estimates.
• Preparing future development workflows for code generation, compilation, debugging, and release support.

See AI Agent Workflow.

Author Role

In the original project, I worked as the project owner / technical lead for the data loop effort, coordinating vehicle-side software, platform backend, platform frontend, algorithm engineering, design reviews, integration testing, and delivery planning.