drawtonomy vs CARLA
Konten ini belum tersedia dalam bahasa Anda.
CARLA is the dominant open-source simulator for autonomous-driving R&D. It pairs an Unreal Engine world with full sensor models — camera, LiDAR, radar, GNSS, IMU, collision, lane invasion — and a Python API for driving the simulation. The companion CARLA Scenario Runner project executes ASAM OpenSCENARIO files against the CARLA world.
What CARLA covers:
- 3D photo-realistic worlds with sensor-grade rendering.
- Camera / LiDAR / radar / GNSS / IMU / collision sensors with ground-truth annotations.
- A Python API for scripting agents, sensors, and traffic.
- OpenSCENARIO 1.x support via Scenario Runner.
- A large ecosystem (Leaderboard, ROS bridge, ML reference agents).
If you are training or evaluating an autonomous-driving stack with sensor input, CARLA is one of the standard tools to reach for.
Licensing
Section titled “Licensing”CARLA’s core is open-source under the MIT License. Some assets and companion projects have their own terms — always check the individual repos.
How drawtonomy relates to CARLA
Section titled “How drawtonomy relates to CARLA”drawtonomy is a free, browser-based whiteboard for driving scenarios. Its in-browser OpenSCENARIO simulator embeds esmini compiled to WebAssembly, so the storyboard runs live in the page — same semantics as native esmini, since the runtime is esmini. (Credit for the OpenSCENARIO 1.x simulation correctness goes to the esmini maintainers.)
drawtonomy and CARLA both simulate driving scenarios, but at very different layers:
| Question | CARLA | drawtonomy (browser simulator) |
|---|---|---|
| Run the OpenSCENARIO 1.x storyboard | ✅ via Scenario Runner | ✅ via embedded esmini-WASM |
| Sensor simulation (camera / LiDAR / radar / GNSS / IMU) | ✅ full | ❌ none |
| 3D photo-realistic world | ✅ Unreal Engine | ❌ 2D canvas only |
| Vehicle dynamics fidelity | high | storyboard-driven (esmini-level) |
| ML / RL training pipeline | ✅ designed for it | ❌ not the target |
| Install required | multi-GB engine + GPU | ❌ runs in any modern browser |
| Time to first run from a fresh laptop | minutes-to-hours | seconds |
| Frame-accurate seek and step on a recorded run | depends on tooling | ✅ scrubber + arrow keys |
.webm export of the simulation | external screen capture | ✅ in-app export, aspect-ratio presets |
| Whiteboard authoring on the same canvas | ❌ | ✅ |
The two tools target different problems:
- CARLA: a sensor simulator. The point is to feed an AV stack realistic perception data and ground truth.
- drawtonomy: a scenario simulator and sketch canvas. The point is to visualise, share, and iterate on the storyboard itself — fast, in a browser, no install.
A reasonable pattern when both are available: CARLA for sensor /
training runs, drawtonomy for a quick browser-side look at the same
.xosc (triage, code review, demos, slides). The .xosc you run in
CARLA Scenario Runner is the same file you can drop onto the
drawtonomy canvas.
When the browser simulator is enough
Section titled “When the browser simulator is enough”- Triage a failing scenario — drop the
.xoscin the browser, scrub to the moment of failure, screenshot. Faster than booting CARLA. - Code review a scenario PR when the reviewer doesn’t have CARLA set up.
- Teach OpenSCENARIO on student laptops without a GPU.
- Record a demo
.webmof the storyboard for a slide, a bug report, or a social post — no screen-capture chain. - Skim a generated catalog from scenariogeneration or Scenic before committing to a CARLA sweep.
When CARLA is the right tool
Section titled “When CARLA is the right tool”- You need sensor data — camera, LiDAR, radar, depth, semantic.
- You’re training or evaluating an AV / ADAS stack.
- You need a photo-realistic 3D world, not a 2D canvas.
- You’re running closed-loop control with a perception stack in the loop.
- You’re integrating with ROS or another robotics middleware.
drawtonomy doesn’t try to compete with any of that. It sits one layer up — the place where you sketch, simulate the storyboard, and review.