autowarefoundation/vision_pilot

What it solves

Vision Pilot provides an open-source, production-ready software stack for Level 2 (L2) Advanced Driver Assistance Systems (ADAS). It enables basic autonomous driving features—such as cruise control and lane keeping—without requiring expensive 3D high-definition (HD) maps, making the technology more accessible for passenger cars, buses, and trucks.

How it works

The system uses a "Hybrid End-to-End AI Architecture" that processes data in parallel using two different AI approaches:

1. Perception AI models for safety.
2. End-to-End AI models for performance.

It relies on a single front-facing monocular camera and integrates three specific AI models: AutoSpeed (object detection), AutoSteer (waypoint detection), and AutoDrive (object presence and road curvature estimation). The system can be run via source build, Debian packages, or Docker containers, and supports integration with the CARLA simulator for testing.

Who it’s for

This project is designed for automotive OEMs, Tier-1 suppliers, and researchers looking to implement or study L2 ADAS features in series production vehicles or logistics transport.

Highlights

• Mapless Operation: Operates in real-time without reliance on 3D HD maps.
• Comprehensive L2 Feature Set: Includes ACC (Adaptive Cruise Control), AEB (Autonomous Emergency Braking), LKAS (Lane Keep Assist), and a single-lane hands-free highway autopilot.
• Low Hardware Requirements: Works with a single monocular camera (1MP-2MP resolution).
• Production-Ready: Designed to be safety-certifiable and available under the Apache 2.0 license.