The history of automotive safety is a history of managing energy. The crumple zone absorbs it. The seatbelt distributes it. The airbag decelerates a body more slowly than the dashboard would. Each of these is a magnificent piece of engineering, and each of them accepts the collision as a given and works on the consequences.
That acceptance was rational. For most of the century there was no instrument that could reliably see a collision developing and no channel to do anything about it in time. Regulation, testing regimes, insurance models and engineering culture all organised around the moment of impact because it was the only moment that could be influenced.
The window was always there
What has changed is not the physics. It is the ability to observe. A collision does not begin at impact; it begins several seconds earlier, when a set of conditions align and nobody has yet noticed. Human drivers notice some of these and miss others. The ones they miss are where the losses concentrate.
Conventional forward collision warning operates in that window, but narrowly — it typically fires when a closing rate crosses a threshold, which is late, and it is looking at one signal. Regulatory minimums for these systems are measured in fractions of a second of advance notice.
An alert that fires at the moment of impact records the event. It does not change it.
Independent controlled testing of fused prediction — road, vehicle and driver state evaluated together rather than separately — has measured alerts arriving three to five seconds earlier than conventional systems, with roughly double the driver reaction time and 99% alert accuracy. Those are the Virginia Tech Transportation Institute figures for our own alert performance, and they are scoped to that testing.
Three to five seconds does not sound like much. At 60mph it is the difference between a collision and a story about a near miss.
The second problem: nobody listens to a system that cries wolf
Earlier detection creates a new failure mode, and it is the one that kills most deployments.
If you look further ahead, you see more potential conflicts, and most of them resolve themselves. A pedestrian steps toward a kerb while you are already braking for a red light. A cyclist appears in a protected lane separated from traffic. A vehicle drifts in an adjacent lane and corrects.
A system that alerts on all of these is technically detecting correctly and practically useless, because within a fortnight the driver has learned to ignore it. Alert fatigue is not a user-experience problem. It is a safety failure with a user-experience explanation.
Which means the interesting engineering question is not “can you see it earlier.” It is “can you see it earlier and stay quiet the rest of the time.” Those two goals pull against each other, and the only way to hold both is to judge severity rather than frequency — to know, from accumulated real-world experience, which developing situations actually end badly.
That judgement cannot be derived from first principles. It has to be learned from a very large number of real situations where the outcome is known.
What this changes downstream
If prevention becomes reliable, several things that currently look fixed start to move.
- Insurance shifts from pricing to preventing. A book of business where collisions fall by half is a different financial instrument, and the loss-prevention effect becomes the product rather than a discount lever.
- Liability moves. When a preventable collision was preventable by a system the operator chose not to deploy, the question about who is responsible changes shape.
- Regulation follows capability. Requirements written around impact survival do not yet know what to do with a system that acts before impact. They will.
- The measurement problem gets harder. You can count crashes. Counting crashes that did not happen requires a comparator, a baseline period and a stated method — which is why every number we publish carries all three.
The part we are least sure about
Prevention is easier to demonstrate in a fleet than in a population. A fleet has a baseline, a defined vehicle set, a comparable period and somebody accountable for the result. Named deployments have reported reductions in at-fault collisions of around 62% and in total collisions of around 68%, each scoped to that customer.
Whether those effects hold at national scale, across mixed vehicles, mixed drivers and mixed roads, is genuinely unresolved. We think they largely do. We do not yet have the evidence to assert it, so we are not going to.