Open-source self-driving stack that skips the HD maps

What it does

Vision Pilot is a C++ and PyTorch-based autonomous driving stack targeting series-production passenger cars, from Level 2 single-lane highway assist up to a planned Level 4 “all road” system. The project releases model architectures, training pipelines, parsing scripts, and weights under Apache 2.0—no black boxes, no vendor lock-in.

The interesting bit

The stack deliberately avoids 3D high-definition maps, running “mapless” or on ordinary 2D sat-nav data. The roadmap also experiments with three flavors of end-to-end AI: component-based (explainable, modular), monolithic (pure neural), and a planned hybrid mode for safety validation. It’s an unusually public attempt to thread the needle between “move fast” neural-network optimism and automotive safety culture.

Key highlights

• Full model weights and training code released, not just inference graphs
• ROS-based runtime with OpenCV and PyTorch under the hood
• Explicit roadmap from L2 lane-keeping → L2+ highway → L2++ urban → L4 full autonomy
• Designed for OEM and Tier-1 integration, not just research fleets
• Apache 2.0 license permits commercial use without patent retaliation

Caveats

• The repo appears to be mostly structural scaffolding and documentation; the actual model implementations and training pipelines are referenced but not deeply shown in the README
• Only “Vision Pilot” (the base L2 version) has a linked subfolder; Plus, PRO, and Drive are roadmap promises without visible code yet
• “Productionizable and safety certifiable” is an ambition; no safety standards (ISO 26262, etc.) are cited in the visible material

Verdict

Worth watching if you’re building autonomous systems and want a fully auditable alternative to Mobileye or Tesla’s closed stacks. Skip it if you need something you can flash to a car this weekend—the hardware integration and mature model code aren’t visible yet.