I love driving, and when I’m out with my son – mostly around suburban Surrey and onto the A- and M- road network – we talk about what the future looks like with self-driving L4 and L5 cars. Will it make journeys quicker? Will we be safer, calmer, more efficient with the bots driving us around? What will the road network look like with enhanced and standardised signage, markings and adapted junctions. It’s nerdy but it’s fun to talk about this stuff and question how a lot of the almost sci-fi might work in practice.
But if you want to understand what L5 autonomy actually means. You can’t start on a motorway. You have to start on a Cornish lane that was ‘designed’ in the reign of Henry VIII by a (let’s assume) man who’s only measurement was ‘avoid straight lines’ and who’s only measurement system was a stick of some agreed length.
My go-to example is the run from a nice Roseland peninsula spot to a tiny fishing village. It’s stunning in the way Cornwall so often is: scenery that makes you consider leaving real life behind – and roads that immediately remind you why you can’t. The lanes are absurdly narrow, bordered by stone hedges that are unyielding fortifications, meeting another car or tractor here is less ‘driving’ than an improvised referendum on who’s reversing today and how much they value their paintwork, tyres and helmsmanship.
This is useful because the autonomy conversation still starts in the wrong places, with the wrong visuals. It starts with wide West Coast roads, clear lines, predictable junctions and a general sense that the built environment is a cooperative partner in the project of motion. In California, “self-driving looks like this” tends to mean something very rational: geometry, visibility, lanes you could land a small plane on and junctions laid out with acres of asphalt and a basic respect for human comprehension.
Naturally, Cornwall is not like this. Cornwall – and plenty of rural Britain – is a place where the road network seems to have been draped over the topography by monks.
The autonomy world has a dull phrase that explains how it expects the world to be: Operational Design Domain (ODD) – the conditions a system is designed to operate within. If you’re keen it’s SAE J3016. The phrase is clinically unsexy, which is part of its charm. It keeps the conversation out of the breathless nonsense that blooms around new technology.
It also implies something people in the industry – notably the people selling you cars – dislike admitting: autonomy is geographical. It exists only where the environment cooperates. And, dear reader, Cornwall does not cooperate, not due to budgets or intransigence – simply because of history.
To illustrate, let’s take a lane standoff: Two human drivers meet head-on and resolve it using social protocols that are not codified. There’s the micro-creep forward, a pause that signals “I’m thinking” or the eye contact that says “I know, this is ridiculous but I’m not reversing into a bramble hedge and a fence post for your Audi Q7”. There’s local knowledge about a passing place behind you that doesn’t look like one unless you’ve been raised on these lanes and have developed the spatial instincts of a maze rat. Then there’s a moral arithmetic too: you’re closer to the wider bit; you’re in the smaller car; you’ve got five people in the back; you look like you’re here for the week and have brought three paddleboards.
Now swap both of these cars for L5 ones and you hit the awkward truth: this route isn’t asking for better navigation, it’s asking for social negotiation. In this context it’s easy to say: the vehicles will simply communicate, agree a plan, one will reverse. The other will wait. Like a pair of Nordic adults having a calm, consensual conversation about responsibility.
Except you’re on a lane with hedges made of rock and 5G signal as powerful as a hamster’s yawn. Even if future communication standards are enhanced, a system that is predicated on connectivity in order to behave sanely in every lane it might find itself in, it’s not “drive anywhere”, it’s “drive where the network hasn’t gone for a lie-down in the sun”.
So assume the vehicles have no helpful network. They must now reason locally. Each one must remember where it has been: the last few hundred metres, where it can reverse, where the lane widens, where sightlines recede, where the verge turns into ‘wet grass hiding a half-metre ditch’. That rolling spatial memory must be table stakes for any serious autonomy.
But memory only gives you so many options. The trouble starts when both vehicles are doing the same thing at the same time, under uncertainty, with no shared authority. If both decide to be cautious (and there are good reasons legally and ethically why they would) you can get a stalemate that looks faintly ridiculous to the human passengers and is, in engineering terms, entirely rational. Two polite robots facing each other, neither wanting to be the first to commit to a manoeuvre that might be unsafe, both waiting for clarity that never arrives. It’s Britain but with fewer apologies and more deadlock.
Now, add the more realistic scenario in the nearer term: one vehicle is automated and the other is driven by a human, a 65-year-old local in a Twingo who’s reversed more cars than you’ve had hot dinners and is perfectly prepared to perform the full choreography of rural right-of-way: gesticulation, eyebrow work, and a hand wave that means “you go”, except it also means “I’m going,” and possibly “you London berk, what are you doing down here?”.
A machine can of course see a hand move. It can classify posture. It can assign probabilities to intent. None of that makes the gesture binding. It must rely only on certainties: where the car is, and what it’s actually doing. The moment you treat informal human signals as authoritative, you’re building a system that can be socially steered into unsafe decisions, which is an inviting prospect for every bored teenager and every slightly unhinged delivery driver in Britain.
So, the automated vehicle does what safety-critical systems do when faced with ambiguity: it becomes conservative. It waits. It yields. It declines to commit until the world becomes clearer and signals replace noise. That’s sensible. It’s also exactly how one ends up with a driverless taxi parked in a single-track lane while the rest of Cornwall ages around it.
There’s another detail we humans grasp instantly: the hidden queue. When you decide to reverse, you’re not only deciding for yourself. You’re making an anxious decision for whatever is behind you, unseen beyond the bend. This is why a simple standoff can turn into a fun village event. A human driver will check mirrors, listen for tyres on gravel, remember the vehicle that turned off behind them a few minutes back and estimate whether this reversal will trigger a chain reaction. A driverless vehicle can’t conjure knowledge of what’s behind the bend unless it can perceive it directly or receive it from sensors elsewhere. To solve this elegantly you need awareness beyond what any single vehicle has. At which point you’re edging toward traffic coordination rather than autonomy: the road as a managed system rather than a stage set for individual decision-making.
This isn’t just a Cornwall problem, rather obviously. Rural Sweden has winter roads with passing places and visibility that vanishes in a sneeze. Japan has mountain villages where lanes are carved by an ancient, determined stream. Rural Ireland is essentially Cornwall with different biscuits. The recurring theme is that it’s the environment that sets the terms. Places built around informal human negotiation are awkward customers for systems that require explicit rules and verifiable intent.
So what happens, what can I tell my son will happen? The answer is mundane. It’s a patchwork. Autonomy that operates confidently on roads that are legible and well-maintained, and then becomes limited, constrained or frankly unavailable in places where the geometry and social protocol defeats it. Sometimes there will be upgrades: defined passing places treated as an asset, clearer priority rules at pinch points, better delineation. Maybe even codified behaviour – a three-flash signal ‘I will yield’, as an example. Sometimes the answer will be operational: remote support when a vehicle gets stuck, or routing that avoids the lane because it cannot guarantee progress. Sometimes the answer will be more bluntly: this road requires a human driver.
None of this betrays the vision, it’s simply what happens when technology meets the world as it actually is: layered, improvised, nuanced and historically messy, full of little social contracts that nobody codified because they didn’t need to.
Cornwall in this sense does us a favour. It stops us pretending and evangelising that autonomy is simple a question of sensors and clever computing. It reminds us that driving is a cultural practice. Some parts of the network are (beautifully) engineered systems. Others are negotiated spaces with hedges and drivers with grudges. Level 5 will thrive where the world behaves like the former. Where it behaves like the latter we’ll either change the road, constrain the autonomy, or just accept that the medieval cart track remains one of the last places a human is still the best interface.
AI: I used AI for the tags, the excerpt, image generation, and a light sub-edit. The ideas, references, observations, and anecdotes are mine.
John Gibbard 