Backblaze Labs
All projects
SampleSample · unsupportedUnsupported · no SLA

CARLA Sensor Data Lake

Turn the open-source CARLA simulator into a synthetic autonomous-driving data lake on Backblaze B2: define scenarios, capture multi-sensor episodes (camera, LiDAR, depth, segmentation, telemetry), and browse and serve the datasets straight from B2.

Built for: Autonomous-vehicle researchers, perception-model engineers, and robotics/sim teams who need scalable synthetic sensor datasets on cheap object storage.

Get the codeCheck project fit

MIT licensed

Dashboard with episode, frame, storage, and scenario metrics and ingest charts
Dashboard — episode, sensor-frame, storage, and scenario counts, a 7-day ingest-throughput chart, and frames-by-sensor / weather / town breakdowns.

What it's built with

Each piece of the stack, and the job it does in this project.

Next.js 16 + React 19
Web UI
Tailwind v4, shadcn/ui, Recharts, TanStack Query
Design system and charts
FastAPI + Pydantic v2 (Python 3.12+)
Backend API
boto3 over the S3-compatible API
B2 storage access
CARLA 0.9.x
Driving simulator (optional, user-hosted)
PyTorch
DataLoader training example (optional)
Backblaze B2
Object storage — the data lake

Core capabilities

What the project does out of the box, before you write any code of your own.

  • Scenarios

    Create, edit, delete, and run reusable CARLA capture configs (town, weather, traffic, sensor rig), stored as JSON in B2.

  • Simulation runner

    Drives the real CARLA server in synchronous mode and streams multi-sensor frames to B2 as they are captured.

  • Episode explorer

    Per-episode metadata and bounding-box annotations with a B2-prefix-scoped sensor-frame browser and presigned-URL preview.

  • Dataset serving

    Presigned-URL frame serving plus a standalone PyTorch DataLoader example that streams straight from B2.

  • Data-lake dashboard

    Episodes, frames, storage footprint, frames-by-sensor, episodes-by-weather/town, and 7-day ingest throughput.

A closer look

More screens from the running project.

Scenarios list of reusable CARLA capture configs
Scenarios — the reusable CARLA capture configs (town, weather, traffic, FPS, frame count, sensor rig) that drive every capture.
Scenario detail showing capture config and captured episodes
Scenario detail — a scenario's full capture config plus the episodes captured from it, with Run / Edit / Delete actions.
Episodes list of captured multi-sensor datasets
Episodes — captured multi-sensor driving datasets in the lake, with scenario, town/weather, frame count, size, source, and status.
Episode detail with metadata, dataset serving, annotations, and sensor-frame browser
Episode detail — per-episode metadata and frames-by-sensor, the PyTorch DataLoader serving command, bounding-box annotations, and a B2-scoped sensor-frame browser.

Choose it for the right job

CARLA Sensor Data Lake is a head start for a specific shape of project, not a supported, general-purpose product.

Use it when

  • You need scalable synthetic multi-sensor driving datasets on cheap object storage.
  • You want to manage, browse, and serve datasets already in B2 from any host — no GPU or CARLA server required.
  • You want to stream frames straight into a PyTorch DataLoader via short-lived presigned URLs.

Choose another path when

  • You need a hosted, multi-tenant SaaS — the API is unauthenticated and bucket-wide by design.
  • You want to capture real episodes without a supported Linux/GPU host running a CARLA server (the carla wheel is Linux/Windows, CPython 3.10 and earlier).
  • You want a managed training pipeline — the PyTorch DataLoader ships as a standalone example, not an orchestrated trainer.
Sample · unsupported

Project status and support

Report defects via the repository's GitHub Issues; this sample is maintained on a best-effort basis and carries no SLA.