Autonomous behavior is often cited as a 5G, low-latency, or edge computing application. That’s a vast oversimplification, in my view, and to understand why, we have to look at the way human reactions take place. That should be the model for autonomous behavior, and it will demonstrate just that could and should be done with low latency or at the edge. A good application to use in our example is autonomous vehicles, since that’s the most-hyped example out there.
Suppose you’re driving down a country road, and a deer steps out onto the road a hundred yards ahead. You see the deer, and you likely take a second or two to see if it’s going to move away. If it crosses into the field or brush, you likely slow a bit in case it reverses course. If it doesn’t move off the road, you then take your foot off the accelerator, and if it still doesn’t move as you get closer, you apply the brakes. This is a fairly typical encounter, and you’d probably do the same thing with a vehicle or person on the road.
Now suppose you’re a hundred feet from the point where the deer (or car, or person) comes out onto the road. You don’t take time to consider here, but instead immediately take your foot off the gas and prepare to break. If you don’t see the subject’s movement off the road very quickly, you apply the brakes. Again, a typical action.
Finally, suppose that you’re on that same road and a deer jumps out 20 feet in front of you. You immediately jump on the brakes aggressively, because that’s what would be needed to avoid a possible collision. Hopefully you don’t have this sort of experience often, but it does happen.
Let’s now try to categorize these three reactions, with the goal of deciding just what and where the reactions are processed.
We could call the first example a reasoned response. There is a trigger, and the trigger sets up an assessment of the situation (when you “took a second or two”). The assessment results in an action, not the initial trigger. After the assessed action, you’d have another assessment, perhaps several, in a kind of loop, until you either pass the point of risk or stop the vehicle.
The second one, we can call a reaction. Here, the trigger stimulates a response, then an assessment of whether the response is appropriate. The response is then assessed as it would be in the first case.
The final case could be called a synapse, which is a direct connection from stimulus to response. There is no assessment until the action, the “panic stop” is complete.
If we want to complete our autonomy framework, we need to add a fourth thing, something completely different, which is a plan. Suppose you’re following a route or heading to a specific destination. You’ll still respond to conditions as in our first three examples, but in addition you’ll be sensitive to other triggers, such as the fact that you’re coming up on a turn or that traffic is getting heavy on your planned route, or perhaps that you’ve been looking for gas or a rest stop, and one is coming up. What we have here is a set of different triggers, things that represent more systemic conditions.
Some 5G proponents will look at this and suggest that all of it is a formula for edge-computing-based, low-latency, applications, but I think we have to put these four things into an autonomous context and validate those claims. To do that, I propose to take them in order of “immediacy”.
The synapse-type response is obviously the one that’s most critical in terms of trigger-to-response latency. In human-response terms, this is the kind of thing where you hope your own reaction time is fast enough. The thing is, we already have on-vehicle automatic braking systems that provide the trigger-response combination. Why would we elect to offload this sort of thing to a network-connected software element, when all manner of things would risk a delay and an accident? In my view, there is absolutely no credibility to the synapse-type response justifying 5G, low-latency connectivity, or edge computing.
The reaction response differs from the synapse in two ways. First, the response to the trigger doesn’t have to be as instantaneous. It’s not instinct to hit the brakes as much as a fast assessment of the situation. Second, the conditions that could give rise to the reaction response are more complex to assess. The deer is 100 feet ahead, and so what specific technology lets us know that it’s a moving obstacle that’s now on the road, or perhaps about to be on the road?
The question here isn’t one of response timing as much as the assessment of the need for a response. A radar or ultrasonic picture warning of proximity is easy to put on-vehicle, but for our reaction scenario, we’d almost surely need to have some form of image analysis. The question is whether the analysis should be on-vehicle where we would have direct access to camera data, or whether it should be remote, in which case we’d have to have real-time network connection to the point of analysis.
I don’t think that autonomous vehicles that demand real-time video from every vehicle is a practical strategy, and certainly not in the near term. Thus, where the visual scene has to be analyzed to provide input into autonomous behavior, the handling should be embedded with the vehicle. Given that doorbells and cameras can be made to recognize faces, eyes, and even animals, I don’t think this is a tall order.
Is it possible that our reactive recognition might be a collaborative function? Could a vehicle system perform an analysis of the scene, and then send the result to a cloud function that would perform an AI analysis on the results? Yes, that would indeed be a nice approach. The relevant frames could be abstracted to focus, for example, on what is moving on or toward the road, and eliminating other distractions. Think of a kind of wire-frame modeling. This could be forwarded to an AI system to compare it to other “incidents” that would allow it to be classified, and the result (an action) returned. The response time doesn’t have to be instant, but it could be a credible 5G and/or edge computing mission.
The reasoned response would be quite similar, and in fact it could be handled by the same kind of logic. All that’s required is that the initial AI assessment return a kind of “look again in x seconds” result, which would then repeat the analysis at that future point. It might also set what could be called a “vigilant” state, where the autonomous system (like a human driver) would be watching more carefully, meaning would be more likely to interpret a condition as requiring a reaction.
Introducing planning changes the dynamic somewhat, but not necessarily as much as might be thought. The reaction time for planned trigger-action combinations can be slower, of course, which makes it practical to do more off-vehicle. The problem is that we already have GPS systems for cars that do most of the planning work for us. Modern ones will get traffic updates and suggest taking alternative routes, too. I think it’s possible that real-time data collection from vehicles could be assembled and aggregated to produce better results than we get today from a GPS, but this isn’t a 5G or edge computing mission; there’s no requirement for sub-second response.
There’s another dimension to autonomous behavior that has to be considered too, and is rarely mentioned. What is the fail-safe procedure? What happens if a self-drive loses its sense of self, if an autonomous big rig barreling down the highway suddenly finds itself unable to operate normally because something broke, or because an essential data connection was broken? We already know that even today’s driver-assist systems, explicitly stated not to be suitable for autonomous operation, result in drivers sleeping at the wheel. We can’t rely on driver intervention or attention, and don’t suggest an alarm to wake the driver. Who knows what their reaction would be?
We need two fail-safe procedures, in fact. One would be targeted at dealing with a system failure of the autonomous element, and the other with some major catastrophic problem that could result in having too many autonomous decisions colliding because nobody knows what others are doing. We’ll take the latter first, because it has an easier answer.
It may be that the strongest argument for central control, even with distributed autonomy as the rule, would be the ability to mediate a response to some catastrophic event. A major accident, a bridge failure, or any number of things that could totally disrupt traffic, could result in synchronized behavior from similar autonomous systems. If everyone turns left at the next intersection to avoid a major traffic accident, the avoidance conditions could become worse than the trigger.
The system failure problem is hard, and there’s no getting away from that. If autonomous systems fail no more often than vehicle mechanics or human drivers, the failures could be statistically acceptable but still enough to create public backlash and even insurance penalties. If they fail more often, then it’s not likely that the technology could survive the bad publicity. The issues could be mitigated if the failure produced a graceful response.
I think that it’s logical to assume that our synapse systems should be fully redundant, and that if they detected a failure of the higher-layer functions, they should flash all the vehicle lights and slowly pull over into a parking lot, the shoulder, or the curb. Obviously sounding a warning to the passengers and/or human driver would also be smart. It would also be smart to have such a fault reported to any centralized traffic or vehicle control function, to facilitate the management of evasion by nearby vehicles.
Where does all this leave us? I think that the majority of the things that 5G or edge computing or low-latency advocates cite to justify their technology of choice in support of autonomous vehicles are nonsense. However, there are clearly things that aren’t cited at all that would ultimately justify 5G and edge computing and latency management. These things, if not handled properly, could threaten the value of autonomous vehicles, not to mention the public at large. If proponents of 5G, low latency, and edge computing are serious, they should look to these issues instead of hyping what’s easy and sensational.