Vision Pilot

Autoware Foundation's open-source L2 ADAS stack from a single camera

License

Apache-2.0

Overview

Vision Pilot is the Autoware Foundation's free and fully open-source L2 ADAS stack, released under Apache-2.0 with the AI model weights included. It runs from a single front-facing monocular camera — 52 to 55 degrees horizontal field of view at 1MP to 2MP — and needs no high-definition maps, following the road in real time instead. The feature set covers adaptive cruise control, forward collision warning, autonomous emergency braking, lane keep assist, lane departure warning, intelligent speed assist and single-lane hands-free highway autopilot. The Foundation positions it as a productionisable, safety-certifiable implementation aimed at automotive OEMs and Tier-1 suppliers, usable commercially as well as for research.

The architecture is what the Foundation calls hybrid end-to-end: camera data is processed in parallel by perception models, which carry the safety case, and by end-to-end models, which carry the performance. Three open-source Autoware models sit underneath — AutoSpeed for closest in-path object detection, AutoSteer for ego-path future waypoint detection, and AutoDrive for end-to-end distance and in-path object presence plus road curvature estimation. Inference runs through ONNX Runtime, and the build is plain CMake with optional ROS 2 interface and an optional occupancy bird's-eye view panel; a Debian package can be produced with cpack. The published spec sheet puts the compute envelope at 3 to 5 INT8 TOPs, and camera support for 8MP at 120 degrees plus front radar fusion is listed as coming.

Why it matters

Driver assistance is one of the few robotics domains where essentially every shipping implementation is proprietary, and where the safety argument is normally the reason given for keeping it closed. Autoware is releasing a complete L2 stack — code and weights — under Apache-2.0 for commercial use, which gives researchers something real to study and smaller suppliers a starting point they could not otherwise build. That it works from one monocular camera with no HD maps also lowers the hardware bar dramatically.

Problem it solves

Production driver-assistance systems are closed, so there is no reference implementation a researcher, a student or a smaller Tier-1 supplier can inspect, evaluate or build on. Vision Pilot supplies one that is intended to be productionisable rather than a demo.

Key takeaways

• Complete L2 ADAS feature set: ACC, FCW, AEB, LKAS, LDW, ISA and single-lane hands-free highway autopilot
• Apache-2.0 including the AI model weights, explicitly permitted for commercial as well as research use
• Runs from one front-facing monocular camera at 52-55 degrees horizontal FoV and 1MP-2MP resolution
• Mapless — no HD maps required, the system follows the road in real time
• Hybrid end-to-end architecture over three open Autoware models: AutoSpeed, AutoSteer and AutoDrive
• ONNX Runtime inference, CMake build with optional ROS 2 interface, and a cpack Debian package target

Who it is for

• Automotive OEMs and Tier-1 suppliers
• ADAS and autonomous driving researchers
• Perception and end-to-end driving model developers
• Transport and logistics fleets exploring bus and truck use cases

Limitations

• L2 only — hands-free autopilot is single-lane, and a human driver must monitor and supervise the system at all times
• The autopilot's stated operating envelope is highways and motorways with clearly marked lanes, in fair weather and visibility, and without construction zones or roadwork
• 'Safety certifiable' describes an intent: certification is a programme an integrator has to run, not something the repository confers
• A single monocular camera means no stereo or lidar redundancy today; 8MP wide-angle and front radar fusion are listed as coming rather than shipped
• Example data and the demonstration video are distributed via Google Drive rather than the repository, so evaluating it requires fetching assets from outside version control