Just wanted to share a really interesting project we came across that shows off what happens when you hook up BeamNG.tech’s physics engine with Autoware’s autonomous driving stack.
Alexander Carballo from Gifu University recently did a presentation showing off this integration, and it’s pretty awesome seeing how our soft-body physics can be used to practically test and train self-driving systems.
A quick breakdown of how they got it working: Basically, they set up a bridge between Windows and Linux using Zenoh. BeamNG.tech runs on Windows, handling all the heavy vehicle dynamics, suspension, and environment simulation. A custom bridge (zenoh_beamng_bridge) shoots the simulation and sensor data over to Linux in real-time. Over on the Linux side, that data gets converted into ROS 2 topics. Autoware reads it, does the actual path planning, and then sends driving commands right back to the car in BeamNG.tech.
A few cool things they’re doing with this setup:
PointCloud Mapping: Pulling LiDAR ray-tracing data straight from the game to map out the virtual environments.
Accel/Brake Calibration: Using our powertrain and tire models to make super precise acceleration and braking maps for the AI to learn from.
Closed-Loop Driving: Letting Autoware fully drive the car while navigating BeamNG’s unpredictable and realistic physics.
It’s always fun seeing the platform pushed in new directions, especially for research where real-world testing is super expensive (and crash-prone!).
Has anyone else been messing around with ROS 2, Autoware, or similar ADAS projects in BeamNG? Drop a comment below, we’d love to see what you’re working on!
Great project showcase! The Autoware integration is definitely impressive.
I’ve been building something similar but at a smaller scale, VisionPilot, a modular autonomous driving system also running in BeamNG.tech. While it’s nowhere near Autoware’s sophistication, the goal is less about replicating a production-grade stack and more about creating an educational, sandbox for experimenting with autonomous systems and the full perception-to-control pipeline.
Currently it has:
Multi-sensor perception: Lane detection (YOLOP + traditional CV), traffic sign/light classification, object detection, sensor fusion with LiDAR/Radar
Visualization: Full Foxglove integration for monitoring
I’m always open to feedback, suggestions, or collaborations, especially from people who’ve played with similar integration challenges or creative testing scenarios in BeamNG.tech.
This is awesome! Really like the focus on making AV systems easier to experiment with and learn from instead of going full production-stack complexity.
YOLOP + sensor fusion + Foxglove in BeamNG.tech is a super fun combo. Would love to see more of the edge-case testing stuff too, looks like a really cool project.
You can watch the showcase here: https://www.youtube.com/watch?v=a9EHYNc7cNQ&t=18s