Safety that can be calculated — not just claimed
Autonomous driving has a fundamental paradox. The more advanced the system is, the fewer mistakes it makes — and the harder it becomes to prove that it is truly safe. The classic approach of "accumulating billions of miles in real-world driving" hits a statistical wall: accidents that happen once in a driver's lifetime almost never appear in test data. Yet it is precisely these extremely rare situations that determine whether an autonomous system will hold up.
Kodiak, a California startup focused on autonomous trucking, addresses this challenge with a methodology called Probabilistic Risk Assessment (PRA). This is not a marketing shortcut, but a thorough engineering tool borrowed from industries where failure is simply not an option: aviation and nuclear energy.
According to Kodiak, PRA combines Bayesian probability theory, systems engineering, reliability analysis, and statistical models into quantifiable results. This allows it to estimate the expected rate of collisions across various severity levels — and, most importantly, to identify the specific scenarios and risk factors that dominate the safety profile.
Three questions that decide between life and death
The PRA methodology breaks down every driving scenario into three fundamental factors:
1. Scenario exposure — how often the autonomous vehicle encounters a given situation
2. Collision probability — what the chance is that a collision will occur in that situation
3. Severity of consequences — how serious the potential consequences would be
A key feature is that PRA explicitly quantifies its own uncertainty. In other words — it can say not only "how big the risk is," but also "how confident we are in this estimate." Where the evidence is weak, Kodiak knows where to focus further testing. No gut feelings, no "it seems safe to us" — only hard numbers.
Kodiak then compares the PRA output against human driver performance benchmarks that the company established in collaboration with leading traffic research centers.
BreakPoint: the AI edge-case hunter
The second pillar of Kodiak's safety strategy is an internally developed tool called BreakPoint. While PRA provides the statistical framework, BreakPoint functions as an AI validator that systematically searches for extremely rare edge cases that could lead to a collision or other undesirable behavior.
BreakPoint does not work randomly. It uses machine learning to intelligently and efficiently hunt down the trickiest situations that real-world road testing might never capture. Think of it as a simulator that deliberately seeks out weaknesses in the system — much like security researchers look for vulnerabilities in software.
The outputs from BreakPoint are then directly fed into PRA models, creating a closed loop: AI hunts for weaknesses, PRA quantifies them, and engineers know exactly what to work on. According to Kodiak, this information flow enables capital-efficient development — the company doesn't need to operate the world's largest fleet; it just needs the smartest data.
Beyond functional safety
The traditional automotive industry relies on functional safety standards to verify safety, which address the question of "what happens when something fails." With autonomous vehicles, however, a more complex problem arises: "Can the system safely handle a real-world situation even when everything is working correctly?"
This is precisely the question addressed by the ISO 21448 standard (Safety of the Intended Functionality, or SOTIF), which deals with risks arising from a properly functioning system encountering unexpected conditions. Kodiak bases its PRA methodology on this standard.
Unlike traditional functional safety analyses, which are performed once and remain static, PRA is, according to Kodiak, a living, iterative process — with every new data point, the models are updated and refined.
From defense to trucking
Kodiak is not a one-dimensional company. The Kodiak Driver technology finds applications in three key sectors: long-haul trucking, the defense sector (autonomous vehicles for the U.S. military), and heavy industry (mining and industrial applications).
In May 2026, Kodiak launched autonomous freight transport in cooperation with carrier Roehl Transport — a practical step from testing to real-world deployment. And it is precisely at the moment when autonomous trucks begin racking up commercial miles that tools like PRA and BreakPoint become absolutely crucial. It is no longer just about demonstrating the technology, but about demonstrable safety to regulators, partners, and the public.
What this means for the Czech Republic and Europe
The European Union does not yet have a unified framework for approving autonomous freight vehicles, but work on one is underway. Kodiak's approach — combining AI-driven fault finding with probabilistic risk modeling — could become an inspiration for European regulatory standards.
For the Czech Republic, as a transit country with a key role in European logistics, the question of autonomous truck safety is highly relevant. Thousands of trucks cross Czech highways every day, and autonomous systems that could eliminate driver fatigue and human error (which, according to U.S. statistics, account for more than 85% of truck accidents) could fundamentally improve traffic safety in our country as well.
For now, however, Kodiak operates primarily in the United States. European deployment of autonomous trucks is a longer-term proposition — but the safety verification methodologies the company is developing will be relevant regardless of which manufacturer first brings an autonomous tractor unit to the D1 highway.
What is the difference between PRA and conventional autonomous vehicle testing?
Conventional testing focuses on accumulating as many real-world miles as possible. PRA, by contrast, uses statistical modeling to estimate safety even in extremely rare scenarios that might never appear in real-world testing. Moreover, it can quantify its own degree of uncertainty — meaning it knows where its estimates are weak and where they are solid.
What is an edge case and why is it so important for autonomous driving?
An edge case is an extremely rare traffic situation — for example, an object falling from a bridge, an animal unexpectedly running onto the road, or an unusual combination of weather and traffic. Human drivers might encounter these situations once in a lifetime, but an autonomous system must handle them safely every time. Kodiak's BreakPoint uses AI to deliberately search for and test these scenarios.
When will we see autonomous trucks on Czech roads?
Currently, autonomous trucks (including Kodiak) operate primarily in the U.S., where the regulatory environment is more favorable. In Europe, pilot projects are underway in several countries, and the EU is working on unified legislation. In the Czech Republic, autonomous truck transport in real operation is not expected before the turn of the decade — and that assumes that safety can be demonstrated using methods such as PRA.
