The Road to Autonomous Driving

The autonomy gap between perception and reality.

A modern vehicle can steer, brake, change lanes, follow navigation and park itself. To a passenger, that can feel close to autonomy. Under the formal responsibility model, however, many of these systems remain driver assistance because the person behind the wheel must continuously supervise the road and take responsibility for the vehicle.

This is the most important distinction in the autonomous-driving debate. The level is not determined by the number of automated features. It is determined by who performs the driving task, who monitors the environment and who must manage failure when the system reaches its limits.

Autonomy is fundamentally a transfer of responsibility – not simply an accumulation of features.

The six levels of driving automation

SAE International defines six levels, numbered 0 through 5. Transport Canada and the U.S. National Highway Traffic Safety Administration use the same framework to explain the division of responsibility between the human and the automated system.

Three systems, three different models

Tesla: sophisticated assistance, active supervision

Tesla Full Self-Driving (Supervised) can navigate routes, steer, change lanes, turn and respond to surrounding traffic. Tesla is equally clear that the current features require active driver supervision and do not make the vehicle autonomous. Under the SAE responsibility model, this places the system in Level 2: the software may control steering and speed, but the human remains responsible for monitoring and intervention.

This point is especially relevant in Canada, where increasingly capable Teslas are common on public roads. Capability can look like autonomy from the passenger seat, but the legal and operational responsibility remains with the driver.

Mercedes-Benz: conditional automation inside a narrow domain

Mercedes-Benz DRIVE PILOT illustrates Level 3. When the system is active within its approved operating conditions, it performs the driving task and the person may turn attention away from continuous road monitoring. The driver must remain available to resume control when requested.

The operating conditions matter: approved road segments, speed limits, traffic conditions, weather and vehicle state define when the system can be used. Mercedes-Benz also emphasizes redundant braking, steering, electrical systems and multiple sensor types. Level 3 is therefore not “hands-free Level 2.” It represents a different allocation of responsibility within a tightly controlled operating domain.

Waymo: no human fallback inside the service area

Waymo operates rider-only autonomous services in multiple U.S. metropolitan areas. Within its approved service areas and operating conditions, the Waymo Driver performs the driving task without a human safety driver. This fits the Level 4 model: the automated system is responsible within a defined operational design domain, even though it cannot drive everywhere under every condition.

This explains an apparent paradox. A robotaxi operating in a mapped service area may be more automated than a privately owned vehicle that can travel almost anywhere. Geographic and operational limits can make higher autonomy achievable sooner because the system can be designed, tested and monitored for a narrower environment.

Vision, radar and lidar: different ways of seeing the road

Autonomous systems must continuously detect the road, vehicles, pedestrians, signs, signals and unusual events, then predict what may happen next and plan a safe response. Companies differ on how much sensing diversity is required.

Tesla has emphasized a camera-led vision strategy supported by neural-network processing. Waymo uses a multi-modal suite combining cameras, imaging radar and lidar. These approaches reflect different assumptions about scale, redundancy, cost and the level of operating-domain control. The market has not yet converged on a single universal architecture.

Why Level 4 may scale before Level 3 becomes ordinary

At first glance, Level 3 should arrive before Level 4 because the number is lower. Commercial deployment is not that linear. Level 3 asks a difficult human-factors question: can a person disengage from monitoring and then reliably retake control when the system reaches a limit? The transition itself can become a safety-critical event.

Level 4 can avoid that handoff by asking the automated system to manage fallback – but only within a narrower environment. A fleet operator can define service boundaries, control maintenance, monitor vehicles remotely, update maps and suspend operations when conditions exceed the system’s capabilities. That operating model is difficult to reproduce across every privately owned vehicle and roadway.

Where Canada stands

Canadian consumers already use many Level 0-2 functions, including automatic emergency braking, adaptive cruise control, lane support and supervised automated driving features. Higher automation is developing through federal safety guidance, provincial rules and controlled testing rather than broad consumer availability.

Transport Canada emphasizes clearly defined automation levels and operational design domains, safe responses when limits are exceeded, cybersecurity, data recording and post-deployment monitoring. Ontario has also introduced a ten-year pilot for approved automated commercial motor vehicles running from 2025 to 2035. The direction is clear: more capable systems are coming, but deployment will be conditional, evidence-based and jurisdiction-specific.

Autonomous driving is an AI governance case study

The road to autonomy demonstrates why advanced AI cannot be governed as a software feature alone. The system combines models, sensors, maps, infrastructure, human behaviour, maintenance, regulation and real-time decisions with physical consequences.

Executive perspective

The closer an AI system moves from recommending an action to taking one, the more explicit its operating boundaries, accountability and fallback controls must become. Autonomous vehicles make that principle visible, but the same lesson applies to AI agents, financial decisions, health care and other high-impact systems.

Key takeaways

• SAE defines six levels of driving automation, from Level 0 through Level 5.
• Levels 0-2 remain human-driven; at Levels 3-5, the automated system performs the driving task when engaged.
• Tesla FSD (Supervised) is not Level 3 autonomy because the driver must continuously supervise and remains responsible.
• Mercedes-Benz demonstrates constrained Level 3, while Waymo demonstrates Level 4 within defined service areas.
• Cameras, radar and lidar provide different information; autonomy depends on the complete system, not a single sensor.
• Progress will depend as much on governance, validation and operating models as on model capability.

Primary sources and further reading