Independent research · Early development

Safely navigating roads without lines.

Production self-driving systems have the same limitation: they won't work without lines. This system, designed from the ground up around automotive safety standards, doesn't.

Millions of miles of paved road are unlined.

Most of these are local suburban & exurban roads which are the last mile of millions of Americans' daily travel. Even when these roads are lined, snow and wear routinely reduce their visibility faster then maintainence can address. Most accidents happen near a drivers' home, so enabling door-to-door autonomous vehicle operation could eliminate a large portion of motor vehicle incidents.

Safety-first design

The bar for autonomous vehicles needs to be safer than the average driver. Research suggests the public expects these systems to be four to five times safer before trusting them. 2 To achieve this while operating effectively and efficiently, safety needs to be the core around which the system is designed & built, not bolted on at the end for a compliance check box.

Current and proposed regulation in the EU and UN, respectively, (The US has no such regulation) require developers to define their own operational design domain (ODD) and necessary safety criteria. 1 Therefore, this project intends to create:

  1. an ODD defined for unmarked and occluded roads
  2. a deterministic safety layer that operates independently of any learned component
  3. an advanced driver assistance system (ADAS)
  4. an autonomous driving system (ADS)
ensuring compliance to ISO 26262 and ISO 21448 from the start.

Safety design

Architecture

Safety
An independent, deterministic layer. Models the swept volume of the vehicle under maximum braking and engages a protective stop if not clear.
Perception
Builds the scene from sensor data and coarse map context: road users, surface geometry, and estimated lane boundaries.
Behavior
Plans a path intended never to trigger the safety layer, with regional traffic law (such as required passing distance for cyclists) enforced.

Full system design

Current Focus

Ground-truth data for unmarked roads is scarce as most development is for highway and urban environments. Collecting sufficient quantities of high-quality data from real life is challenging, so simulation-first development is being pursued. CARLA is being utilized for this but environment modeling and scenario creation are necessary. In parallel, most of the effort has gone into the regulatory groundwork: defining an ODD for unmarked roads, specifying how to calculate where the vehicle will travel, and shaping a safety model that does not trip during normal operation.

References

  1. EU Implementing Regulation 2022/1426, European Union.
  2. Liu P, Wang L, Vincent C. Self-driving vehicles against human drivers: equal safety is far from enough. J Exp Psychol Appl. 2020;26(4):692–704.