For almost two decades, autonomous driving has existed in a strange place between science fiction and reality. Every year, millions of people entrust their lives to software-assisted driving systems, and increasingly those systems are making decisions that once belonged exclusively to human drivers. Despite headlines predicting fully self-driving cars "next year," the industry has learned that autonomy will not be achieved through a single breakthrough, but through thousands of engineering problems being solved one by one.
Nevertheless, meaningful progress is undeniable. Millions of vehicles now incorporate advanced driver-assistance systems (ADAS), and commercial robotaxi services are operating without drivers in selected environments. The question is no longer whether vehicles can drive themselves in certain conditions. The question is how these systems work, and how engineers can prove they work safely.
In many ways, autonomous driving is ultimately a system engineering and validation problem, and the industry's long-term success will depend on demonstrating that these systems can behave safely, predictably, and repeatedly under real-world conditions.
Key Takeaways:
- The SAE defines autonomous driving across six levels, from Level 0 (fully manual) to Level 5 (fully autonomous).
- "Self-driving" and "autonomous" are not interchangeable. A self-driving car still expects a human ready to take over; a fully autonomous car does not.
- The three fundamental components of any autonomous system are perception, planning, and control.
- Most vehicles perceive the world around them through a sensor suite consisting of various cameras, radars, LiDARs, and ultrasonic sensors, coordinated and arbitrated through sensor fusion.
What is Autonomous Driving?
An autonomous vehicle senses its surroundings and drives itself without human input. It relies on sensors, processors, and control software to perceive the road and decide what to do, performing tasks an attentive human driver would.
The terminology is looser in everyday use than in engineering. The SAE (Society of Automotive Engineers), whose framework is the industry reference, prefers the word automated to autonomous, since true autonomy implies a kind of decision-making that reaches beyond the electromechanical scope of a car.
The difference between self-driving and autonomous matters just as much. A self-driving car can handle some or all the driving but still assumes a human is present and ready to take over. A fully autonomous car assumes no one is. That single distinction is what separates a driver-assist feature from a vehicle that needs no driver at all, and the six SAE levels exist to make it precise.
What are the Levels of Autonomous Driving?
The SAE defines six levels of driving automation, numbered 0 to 5. The line that matters most falls between Level 2 and Level 3. At Levels 0 through 2, a human is driving and the system assists. From Level 3 upward, the system does the driving and the human steps back, progressively, until Level 5 removes the driver entirely.
| Level | Name | Who drives when engaged | Human role |
| 0 | No Driving Automation | Human | Performs all driving; system limited to momentary warnings or interventions |
| 1 | Driver Assistance | Human, with one assist | System handles either steering or speed, never both at once |
| 2 | Partial Driving Automation | Human, supervising | System handles steering and speed together; human must watch constantly |
| 3 | Conditional driving Automation | System, in set conditions | Human driver must take over when the system requests |
| 4 | High Driving Automation | System, within a defined zone | No takeover needed inside that zone |
| 5 | Full Driving Automation | System, everywhere | None; the vehicle drives in any condition a human could |
Level 0: No Driving Automation
The human controls all driving. Features such as automatic emergency braking or blind-spot warnings may step in for a moment, but they do not drive the car.
Level 1: Driver Assistance
The system manages a single continuous function, either steering or speed. Adaptive cruise control and lane-keeping assist are the common examples. The driver stays responsible for everything else.
Level 2: Partial Driving Automation
The system controls steering and speed at the same time, for instance pairing adaptive cruise control with lane centering. The driver must keep watching the road and be ready to act at any moment. Most features marketed as "autonomous" today, including Tesla Autopilot, Ford BlueCruise, and GM Super Cruise, sit at this level.
Level 3: Conditional Driving Automation
The system handles the full driving task within specific conditions, such as slow highway traffic, and monitors the environment itself. The driver can disengage but must be ready to retake control when prompted.
Level 4: High Driving Automation
The system drives without human intervention inside a defined operational domain, such as a mapped urban area or a set of conditions. Waymo and Zoox's robotaxis operate here.
Engineers refer to these boundaries as the vehicle's Operational Design Domain (ODD)—the specific roads, weather conditions, speeds, and operating scenarios for which an autonomous system has been designed, tested, and validated. Expanding a vehicle's ODD is one of the central engineering challenges in autonomous driving.
Outside its validated ODD, the vehicle will not drive itself.
Level 5: Full Driving Automation
The system drives anywhere a human could, in any condition, with no expectation of human involvement. No vehicle on the market has reached this level.
How Does Autonomous Driving Technology Work?
Autonomous driving control systems work in three functional stages that mirror how a person drives.
- The first is perception. The vehicle gathers data about its surroundings through an array of sensors, building a live model of the road, the traffic around it, pedestrians, and signage. The resulting data must be processed in milliseconds by specialized computing hardware before the vehicle can decide and act.
- The second is planning, the system's equivalent of judgment, where software weighs the perception data against high-definition maps to calculate a safe route and the commands needed to follow it.
- The third is control, where the control system translates those commands into steering and speed changes through drive-by-wire actuation
Perception is generally regarded as the hardest of the three, because a car can only react to what it manages to detect. That is why sensor choice is so vigorously debated in the field, and it is where the next section goes into detail.
Sensors Used in Autonomous Vehicles
Autonomous vehicles build their perception of their surroundings from four main sensor types: cameras, radars, LiDARs, and ultrasonic sensors. Each one reads the environment differently, and each has one or more weaknesses that can be compensated for by the strengths of the others. This is why most self-driving systems do not rely on a single sensor but combine several, a method known as sensor fusion.
The sections below cover what each sensor does well, where it falls short, and why the combination matters more than any individual component. For a head-to-head technical comparison of the three primary perception sensors, we have a dedicated breakdown about radar vs LiDAR vs camera.
Cameras
Cameras are the most mature and least expensive sensing technology on the road. They are small enough to sit discreetly behind the windshield or in the bodywork, and they are the only sensor that reads color and interprets text, which makes them indispensable for recognizing traffic lights and road signs. In effect, they replicate what a human driver does with their eyes.
That strength is also their constraint. A camera sees only what is in front of it, so a blocked lens or poor lighting degrades its input the same way it would for a human. Cameras also produce raw images with no inherent sense of distance, so they depend entirely on machine learning models and a processor capable of interpreting the scene in real time. Averna vision and optical inspection expertise sits squarely in this domain.
Radar
Radar emits radio waves and measures their reflection, which lets it calculate both the distance and the speed of objects around the vehicle. Its main advantage is reliability in conditions that defeat optical sensors. Rain, fog, dust, and darkness have little effect on radio waves, and radar holds its accuracy at long range, which makes it well suited to highway speeds and adaptive cruise control.
The trade-off is resolution. Radar can tell that an object is present and how fast it is moving, but it produces a coarse picture that cannot read a sign or reliably distinguish a pedestrian from a lamppost. Newer imaging radars narrow this gap by adding height data to what radar already measures.
LiDAR
LiDAR (light detection and ranging) emits laser pulses and measures the time each pulse takes to return, producing a precise three-dimensional point cloud of the surroundings. Where a camera infers depth and radar approximates shape, LiDAR measures both directly, mapping the exact position and contour of objects in fine detail. It is unaffected by lighting conditions, so a bright light or a dark road does not distort its reading.
Its limitations are cost and physical size, along with a sensitivity to weather. LiDAR units have historically been expensive and bulkier than cameras or radar, though prices have fallen as the technology matures. Fog, snow, and heavy rain scatter the laser and reduce accuracy, and like radar, LiDAR cannot read color or text. Waymo (owned by Google’s Alphabet) is the highest-profile adopter, building its perception stack around LiDAR rather than betting on cameras alone.
Ultrasonic sensors
Ultrasonic sensors emit high-frequency sound waves and measure their echo to detect nearby objects. They operate only at short range, which makes them the workhorse of low-speed maneuvers such as parking and close-quarters obstacle warnings. They are inexpensive and well understood, which is why they have been standard on production cars for years.
Their range and resolution are too limited for high-speed driving decisions, and their performance can drop in certain weather, so their role is supporting rather than central.
Sensor Fusion: How These Sensors Work Together
No single sensor delivers a complete and reliable view of the road.
- Cameras read context but not depth
- Radar measures motion but not detail
- LiDAR maps geometry but not color
- Ultrasonic sensors cover only the immediate surroundings
No sensor is universally superior. Every perception architecture represents a series of engineering trade-offs involving performance, cost, packaging, environmental robustness, computational complexity, and functional safety requirements.
Sensor fusion is the process of merging these complementary inputs into one coherent model of the environment, so that the weakness of one sensor is offset by the strength of another.
Different manufacturers balance these tradeoffs differently. Some rely heavily on camera-based perception, while others combine multiple sensing modalities to provide complementary information and redundancy.
The industry remains divided on the optimal sensor strategy. Most developers employ multiple sensing modalities to provide complementary information. We’ve summarized the ideas above in a table.
|
Sensor |
What it detects |
Strengths |
Limitations |
|
Camera |
Color, text, objects |
Mature, low cost, reads signs and signals |
No native depth, heavy processing demand, struggles in glare and low light |
|
Radar |
Distance and speed of objects |
Reliable in rain, fog, and darkness; long range |
Low resolution, weak object classification |
|
LiDAR |
3D shape and exact distance |
High precision, lighting-independent |
Costly, bulkier, degraded by fog and snow, no color or text |
|
Ultrasonic |
Close-range obstacles |
Inexpensive, proven, accurate up close |
Very short range only, weather-sensitive |
ADAS VS Autonomous Driving: What's the Difference?
The difference comes down to who is responsible for driving. ADAS, or advanced driver-assistance systems, support a human who stays in charge of the vehicle. Autonomous driving replaces that human, at least within defined limits. In SAE terms, everything through Level 2 is ADAS, and Level 3 upward is autonomous driving.
The confusion is understandable, because the features can look alike from the driver's seat. Adaptive cruise control, lane centering, and self-parking are all ADAS. They handle parts of the driving task but assume a human is watching and ready to act. A Level 2 system that steers and controls speed at the same time is still ADAS, however capable it feels. The moment the system takes full responsibility, and the human is no longer required to supervise, the vehicle has crossed into autonomous territory.
Validating an assistance feature and validating an autonomous system are two different problems. The first proves a feature helps a human; the second proves a system can safely replace one, which is where ADAS testing and validation becomes its own discipline.
What are the Benefits of Autonomous Driving?
The case for autonomous driving rests first on safety. Most crashes trace back to human factors such as distraction or fatigue, and a machine does not get tired or check its phone. The US Department of Transportation attributes 94% of crashes to human error in its National Motor Vehicle Crash Causation Survey, though NHTSA's own source report cautioned against reading that number as the cause of crashes, and the claim remains contested. The stronger evidence is empirical: Waymo, the largest operator, reports a 90% reduction in serious-injury crashes across more than 127 million fully autonomous miles, measured against human drivers on the same roads.
Beyond safety, the benefits center on access and time. Autonomous vehicles can extend mobility to people who cannot drive, including many elderly and disabled travelers. They can hand commuting hours back to passengers as usable time. At scale, coordinated autonomous traffic could ease congestion and lower emissions through steadier driving, although these system-level gains depend on adoption levels that do not yet exist and remain projections rather than measured outcomes.
Beyond passenger vehicles, these same technologies are being deployed in logistics, mining, agriculture, construction, and other industries where repetitive operational or driving tasks can be automated under controlled conditions.
Why Autonomous Driving Is Harder Than It Looks
Humans drive using a remarkable amount of contextual reasoning. We recognize unusual situations, infer intent from subtle cues, and adapt instantly to conditions we have never encountered before.
Autonomous systems must accomplish these same tasks through sensors, algorithms, and computing hardware operating under strict safety constraints. A child chasing a ball, temporary construction signage, glare from the sun, or a partially obscured lane marking can become a difficult perception problem.
The challenge of autonomous driving is not simply making a car drive itself. The challenge is ensuring it behaves safely during the rare situations that occur only once in millions of miles of operation. Engineers often call these rare situations "edge cases"—events that occur infrequently but may expose weaknesses in perception, decision-making, or control algorithms.
Is Fully Autonomous Driving Possible?
Partly. The honest answer in 2026 is that autonomous driving works within limits, not everywhere. Level 4 vehicles operate commercially today, but only inside mapped, geofenced areas under conditions they are designed to handle. Waymo runs driverless services across roughly eleven US cities, yet it built that capability largely in Sun Belt locations chosen for their wide roads and mild weather.
Full Level 5, a car that drives anywhere a human could in any condition, has not been achieved by anyone, and much of the research community expects it to stay out of reach for the foreseeable future.
Moving from a controlled environment to the real world is the main challenge that remains, and it is above all a validation issue.
Leaders in Autonomous Driving
With little argument, Waymo leads. The Alphabet subsidiary has completed more than 20 million autonomous rides, runs around 3,000 vehicles at Level 4 across roughly eleven US cities. Its first markets outside the United States, London and then Tokyo, are planned for late 2026.
Source: Waymo.
The closest global challenger is Baidu, whose Apollo Go service describes itself as the world's largest autonomous operator by ride volume, with more than 17 million cumulative trips across 22 cities and a European expansion underway. Tesla takes a different path. Its robotaxi service operates in Austin and the San Francisco Bay Area, but with safety supervisors on board rather than fully driverless, and its consumer cars remain Level 2. Amazon's Zoox, building a purpose-designed robotaxi, is moving toward paid service in 2026, while Chinese operators WeRide and Pony.ai run commercial fleets and are expanding abroad.
Cruise was once seen as a frontrunner, but GM shut down its robotaxi unit in 2024. However, they have now shifted their focus to Super Cruise, their hands-free driver assistance system.
How are Autonomous Driving Systems Tested and Validated?
Every claim in this article, from sensor reliability to Level 4 safety, ultimately rests on testing. This is the hardest and least visible part of autonomous driving, because a system that works 99% of the time is not good enough when the remaining 1% involves human lives. Validation must cover the edge cases that almost never happen, across every sensor and every condition a vehicle might meet.
Testing Sensor Fusion Systems
Two things make it especially demanding. First, sensor fusion must be verified as a system rather than sensor by sensor, since the dangerous failure modes appear precisely when cameras, radar, and LiDAR disagree.
Sensor fusion is often described as redundancy, but it is better understood as complementary information. Cameras understand context. Radar measures motion extremely well. LiDAR provides precise geometry. By combining these data streams, engineers can improve confidence in the vehicle's understanding of its environment.
The challenge is that fusion itself becomes another system requiring validation. Engineers must understand not only how each sensor performs individually, but also how disagreements between sensors are detected and resolved.
Validation is arguably the defining challenge of autonomous driving.
Traditional automotive testing often verifies that a component performs as specified. Autonomous systems require something much more demanding: demonstrating acceptable behavior across an effectively infinite combination of roads, weather conditions, traffic scenarios, and human interactions.
Simulation and Hardware-in-the-Loop Testing
Engineers therefore rely on a hierarchy of validation techniques. Consider a typical validation pyramid, starting from the base:
- Component characterization of sensors and ECUs
- XIL and Hardware-in-the-Loop testingof sensors and ECUs
- Simulation and scenario generation
- Closed-loop system validation
- Controlled proving grounds
- Real-world fleet deployment
No organization can physically drive enough miles to encounter every conceivable edge case. As a result, virtualization, simulation, and automated test methodologies have become essential enablers of autonomous vehicle development. This need for scalable, repeatable validation has driven the development of advanced XIL platforms, radar target simulators, sensor simulation technologies, and automated test environments capable of reproducing rare scenarios on demand. Averna's Sensor Validation Testers and Sensor Fusion XIL solutions are examples of technologies built specifically for these challenges.
Functional Safety and SOTIF
Autonomous systems are expected not merely to function but to fail safely. This expectation has elevated standards such as ISO 26262 and ISO/PAS 21448 (Safety of the Intended Functionality, or SOTIF) into central design considerations.
ISO 26262 primarily addresses failures caused by faults within electrical and electronic systems, while SOTIF addresses hazards that can arise even when systems operate as designed, such as sensor limitations or ambiguous environmental conditions.
Validation activities increasingly focus on demonstrating not only that systems work correctly, but also that they respond appropriately when sensors become degraded, inputs become ambiguous, or operating assumptions are violated.
Autonomous driving will almost certainly arrive incrementally rather than all at once. Some systems will rely primarily on vision, others on sensor fusion. Some will remain confined to specific operating domains, while others will gradually expand their capabilities.
The Future of Autonomous Driving Relies on Validation
Regardless of the architectural approach, one principle remains constant: trust in autonomous systems is earned through rigorous testing and validation. The future of autonomous driving depends not only on better sensors and smarter algorithms, but on demonstrating that these technologies can operate safely, predictably, and repeatedly under real-world conditions.
No single sensor technology, software architecture, or AI model is likely to define the future of autonomous driving. Different companies will continue to pursue different technical approaches, each with its own strengths and tradeoffs. What they all have in common is the need to validate increasingly complex systems across an almost limitless range of real-world scenarios. That challenge has driven significant investment in perception validation, sensor simulation, sensor-fusion validation, and other advanced test technologies. Ultimately, the pace of autonomous driving adoption will depend not only on innovation, but on the industry's ability to prove that these systems deserve the public's trust.
To explore a more optimized approach to ADAS / AD testing for your own program, speak with one of Averna's experts.
You may also be interested in…
Autonomous driving wouldn’t get anywhere (pun intended) without radar! See how we helped this automotive giant accelerate manufacturing in this case study.
