By Dr. Lance Eliot, the AI Trends Insider
In my college days, we used to drive from Los Angeles up to the Bay Area so that we could watch our football team play against several of our most bitter rivals. The driving time was about six to seven hours and was quite a journey. Me and my college buddies would pile into several cars and act like a caravan for the driving journey. Whomever was at the front of the pack of cars would let the rest of us know whether there was anything up ahead to be mindful of.
The ad hoc news provided to us by the lead car included aspects of the utmost importance, such as a speed trap being maintained by the California Highway Patrol up ahead (we certainly didn’t want to get caught speeding!), and encompassed rather mundane aspects such as a herd of cows a few miles in front of us that we’d be able to see off to the right in some grazing pasture. Notably, sometimes the updates were quite crucial safety tips – I remember one time that the lead car saw a couch fall off the back of a pick-up truck and alerted the rest of us to be on the watch for it, as it was sitting in the middle lanes of the highway and could have caused any of us to get into a dreadful car wreck.
We’d usually start the journey and be just a few car lengths away from each other, therefore the news items weren’t much of a heads-up since we could see what the cars ahead of us saw. But, after about an hour of driving, we’d all invariably get spread out. The lead car might be a few miles ahead, and the last car in the caravan might be several miles behind the next-to-last car in the sequence. You’d maybe think that only the lead car would have something useful or significant to say, but that’s not only the case. One time, the trailing car notified the rest of us that a motorcade of police was barreling along on the highway at a very high speed and rushing past him and would likely catch-up with the rest of us and the lead car in maybe fifteen minutes at the speed they were going.
I admit that this was in an era before we had GPS readily available and there wasn’t Google Maps and nor specialized apps like Waze. Today, those kinds of handy tools provide helpful traffic related information and are often essential for navigating the roadways. Those tools do a good job of providing insights about the roads, but they still aren’t quite fully all-encompassing. If you are driving on a road that has a tight curve up ahead, none of those apps will necessarily in real-time at the moment that you need to know be able to tell you that there’s a car that just stalled there about thirty seconds ago, and if you take the curve tightly you might ram into it. Only once you’ve started into the curve will you realize what’s afoot, and hopefully have the driving skills to avert a dangerous situation.
What does this have to do with AI self-driving cars?
At the Cybernetic Self-Driving Car Institute, we are working on helping to make self-driving cars more savvy about roadway conditions in real-time so that they will be better prepared during actual journeys on the roadways.
Some have wondered whether AI self-driving cars will be able to tap into GPS, Google Maps, Waze, and other such tools to help navigate the roadways. Yes, by-and-large, most of the auto makers and tech firms that are developing self-driving cars are making provisions to utilize those kinds of tools. That’s relatively straightforward and there’s not much trickery or special capabilities needed to tap into those info sources and use them.
One step even further though involves trying to make AI self-driving cars essentially omnipresent.
Now, I realize that this is not the divine kind of omnipresence, and some are maybe even a bit offended at the use of the word omnipresent being used in this context. Excuse the use of the word, and please go along with the overall meaning or spirit of what the word entails. Allow me a moment to explain.
The notion is that self-driving cars will be able to communicate with each other, and too the roadway infrastructure, doing so in a manner that they will each be able to warn or inform the others about real-time roadway conditions.
In a sense, it’s kind of like how me and my buddies would provide updates during our caravan trips. A self-driving car ahead of my self-driving car might communicate to my self-driving car that it has just observed a stalled car at the curve ahead of me, which it already passed through, and so now my AI will be forewarned as to the stalled car. The AI of my self-driving car will presumably change how it was going to take the curve, since it now knows that the stalled car might well be stuck there and could get rammed into.
There is a large push toward V2V (vehicle-to-vehicle) communications going on in the car industry right now. You can do V2V to any kind of car, whether a human driven car or a self-driving car, though certainly the self-driving car to another self-driving car is going to be the most seamless way of communicating about driving conditions. Similar to my earlier point, the AI of a self-driving car ahead of you might via V2V communicate to the AI of your self-driving car about a vehicle that is stalled on the curve ahead. Or, maybe it warns you about a couch that’s sitting in the middle of the highway ahead of you. And so on.
Besides V2V, there is also V2I (vehicle-to-infrastructure), which involves cars that communicate with the roadway infrastructure. For example, a bridge up ahead of me might have one of the lanes closed due to an accident that happened on the bridge, and so the bridge itself communicates to my car to let me know about the situation. The other day there was a flooded street near where I live, and I only found out once I turned onto the street and there was an electronic board sign displaying a message to turn back. It would have been more helpful if the electronic board sign was broadcasting an electronic message to nearby cars to let them know well in-advance of getting onto the street. This would have avoided having tons of cars that were all trying frantically to make U-turns.
There’s also V2P, which is somewhat controversial, and refers to vehicle-to-pedestrian communication. This would allow pedestrians to communicate out to cars, and likewise for cars to communicate to pedestrians. For example, the other day there was a school that was walking several classes over to a nearby store, and the young students were all trying to cross the streets. A teacher that was guiding the students could have used a pedestrian electronic link to let nearby cars know that children were crossing the street and to therefore drive with extra caution.
Some argue that having pedestrians communicating with cars is fraught with issues. Suppose a pedestrian tries to trick cars into not driving on a particular street and so the pedestrian sends a message that a six-hundred-pound gorilla is standing in the middle of that particular street (well, maybe not a gorilla, but you know what I mean, someone making a false claim to discourage car traffic).
This does bring up the whole topic of trustworthiness in any of these kinds of communications, whether it be V2V, V2I, or V2P (the collective set of such communications is called V2X). Just because a car up ahead of you says there is a couch in the middle of the roadway, why should you believe it? Maybe the car up ahead is trying to trick other cars. Or, maybe the car up ahead truly believes there is a couch there, but there isn’t one. For the V2I, we would likely hope that the infrastructure would be more trustworthy since it likely is being maintained by a governmental agency, rather than a car-to-car communication which might be coming from just anybody.
Another important factor is the timeliness of the communications. If a V2V takes place, and the self-driving car ahead of you is let’s say completely truthful, and it says that there is a stalled car around the curve, you need to weigh the timing of the communication. Suppose the self-driving car that’s ahead of you had sent that message just now, but that it was based on having driven through that the curve twenty minutes ago. Perhaps by now the stalled car is no longer there. Your self-driving car might be using information that is outdated, and therefore take actions that are no longer needed or appropriate.
On the topic of timeliness, when I drove down that flooded street later that night, I noticed that the electronic sign board was still saying that the street was flooded. But, by nighttime, the waters had receded, and the street was just fine. Obviously, the street maintenance crew had left the electronic board sign there and had neglected to change the message or remove the sign itself. This is another example of the timeliness of information, and in this case imagine if the V2I was broadcast to cars that the street was flooded. Thus, even if we consider V2I to potentially be more trustworthy, it’s going to be vulnerable to lack of timeliness and other undermining factors.
So, we want self-driving cars to know what’s around the next corner and beyond its horizon, but we also know that getting such information can be difficult, plus knowing whether the information is timely and accurate is definitely a concern. If we’re aiming for some kind of true omnipresence, the self-driving car and its AI have to be aware of the real-world and not fall for fakery, ill will, or other adverse aspects.
As an example of how extreme things could get, imagine a criminal that had robbed a bank and wants to have a clear path for a getaway. They could have their self-driving car tell other nearby cars that traffic is clogged up and those cars should avoid the congestion by going down a different street. This could open the roadway for the getaway car. And, if done really cleverly, maybe even block the police that might be coming down that other street into which traffic is now heading. I realize this seems like a plot in a bad science fiction movie, but I am just pointing out that we’ll need to have safeguards that take these kinds of aspects into account.
How can the AI of the self-driving car know whom to trust?
One viewpoint is that the AI should be weighing the sources of the data and give more or less weight to various sources. The V2I might have a high weighting, while the V2V might have a lower weighting since it is considered less trustworthy. If the V2V is coming from a police car or official government vehicle, perhaps the weighting goes to high.
Another aspect involves comparing multiple sources. If there are a lot of V2V’s happening all at once, and if the preponderance of them are saying there’s a couch in the middle of the highway, this would be considered more likely as being a real report versus if only one other car made such a claim. The idea is to use crowdsourcing as a method and via voting or other conflict measurements try to ascertain what’s true versus what is not.
This brings up another facet of these communications, namely how many of them can your self-driving car handle at once? Suppose that there are a hundred other self-driving cars around you, and all of them are bombarding your AI with what’s going on. Second by second, or maybe even much faster, such as every millisecond. Meanwhile, let’s suppose that the roadway infrastructure has dozens of broadcasting sites near your self-driving car as it is traveling along on the highway. How does your self-driving car decide which of these sources to give attention to? Trying to decipher all of them at once might be daunting and consume a tremendous amount of on-board processing and memory. Right now, we have so few V2X’s in place that it is an easy task for an experimental self-driving car to cope with, but once we have lots of V2V, V2I, and V2P, it will be a deluge of data, some of which is useful and some not, some of which is timely and some not, etc.
Notice that I’ve been describing the communications as though the information conveyed is crisply stated and summarized, but it doesn’t have to be that way. The data could be raw data coming from another self-driving car’s cameras or radar or other sensors. In that case, the self-driving car that receives the data would be able to analyze and interpret it, rather than relying upon the interpretation provided by the other self-driving car. The downside with interpreting it would include the time needed by the receiving self-driving car to do so, plus it might not be familiar with the type of camera images or radar data that is being provided (the other self-driving car might be using a different make or model than the sensors on the receiving self-driving car).
This brings up the aspect that we’ll need some universal protocols for the V2V, V2I, V2P, which are indeed being formulated, along with how to interpret the data. In addition, the data needs to be timestamped so that if there’s a delay during communicating, the receiving self-driving cars will know that the data is time lagged.
The time lag could be crucial in life-or-death circumstances. If the self-driving car ahead of me has just rammed into another car, and if my self-driving car has not detected the crash, and meanwhile the self-driving car involved in the crash has sent out a message, we need to consider how long will it take for the message to be emitted and received. Suppose the V2V is occurring by use of the Internet, and so the self-driving car in the crash sends out a message across the Internet, and assuming that there is a link, and assuming that my self-driving car also has a connection to the Internet, it could take a lot longer than the time it takes for my car that’s going 80 miles per hour to pile into the crashing cars.
It would also likely be beneficial to classify messages that are going to be sent back-and-forth. A message that offers a danger ahead warning of an impending accident scene would seem to be more important than a message that says the road ahead has some minor potholes. Messages should be conveyed as to whether the indications are of an immediate concern or an overall concern, and how much priority should be given to the message.
Another question involves whether all self-driving cars are going to be required to send and receive messages, or whether you can opt out of having your self-driving car participate. Maybe you don’t want your self-driving car to send messages, and only receive messages. Maybe you only want to send messages, but not receive messages. Or, perhaps you don’t think any kind of messaging should be undertaken by your self-driving car. Will we as a society require that messaging occur or allow it to be voluntary?
Furthermore, will the messages be attributable to a vehicle? In other words, maybe the messages should be sent anonymously to allow for privacy. But, some would say that anonymous messages would lend itself to people sending trick messages and not taking things seriously. There are those that believe that your self-driving car should be sending messages and should also be saying that it is the one sending the messages, thus allowing for being able to trace where the messages came from. We could use some kind of specialized ID that your self-driving car emits but is only known as being your self-driving car by getting a court order. These kinds of arrangements though have various downsides and trade-offs.
Some believe that our self-driving cars will be able to build-up trustworthiness over time. Let’s pretend that your self-driving car has reported one hundred valuable roadway aspects, and so it is considered a more trustworthy source of roadway info. Some say that we might even setup micropayments, whereby self-driving cars that the owners opt to have participate in messaging will get paid to do so. Or, maybe it will be ad based. Or, some think the government should compensate those that allow their self-driving cars to aid in traffic management.
If you’re interested in the topic of data exchange between vehicles and also roadside infrastructure, you ought to take a look at the Wireless Access in Vehicular Environments (WAVE) standard. The standard and the IEEE 802.11p are part of the set of various protocols that are being developed and fielded for these V2X purposes. This also includes VANETS, Vehicular Ad-hoc Networks, and the use of DSRC’s (Dedicated Short-Range Communications). The infrastructure would have RSU’s, RoadSide Units, which are doing the broadcasting.
Some vendors and developers are pushing forward on the omnipresence goal. For example, an Australian company called Cohda has been developing systems for “surround vision” using Nvidia’s Drive platform. Their hope is to further provide 360-degree awareness for AI self-driving cars, allowing any self-driving car to essentially see around the bend by collecting and synthesizing data from other nearby self-driving cars. With the advent of 5G mobile networks, we’ll gradually have the needed speed to try and ensure that the communications are happening on a timely enough basis.
As mentioned, some of the most limiting factors involve connectivity issues, bandwidth constraints, signal fading, routing aspects, and the like. Furthermore, there are some quite serious privacy issues that arise, along with crucial security considerations. Imagine if a terrorist were able to break into the omnipresence and broadcast messages that misled self-driving cars into taking dangerous actions.
One last aspect to consider involves whether the communications should also include advice from the communicating source. Suppose a self-driving car up ahead reports that there’s a couch in the middle of the roadway. This is informative to my self-driving car, which then the AI decides what to do. If the self-driving car ahead had also suggested that other cars behind it should get off the highway, would this be helpful or not to the receiving self-driving cars. It might be helpful in that maybe that’s the right thing to do. Or, it might be misleading or even distracting and not be a prudent step to take. It’s one thing to say what is occurring or what is detected, and its another to then make a recommendation or stated action for others to take.
In whatever manner this all plays out, there’s no doubting that we’ll ultimately have V2V, V2I, V2P, and the technology will be developed sufficiently to allow it to occur. The rules of how it is to be used, and what it is to be used for, will require not just a technological perspective, but also require a societal, political, business, and ethical perspective. We’re still in the infancy of these capabilities and it’s timely to start laying a foundation for what this will become. Omnipresence should be aimed at being a good thing, and not be allowed to fall into something that is ill-used and that endangers our roadway safety.
This content is orignally posted on AI Trends.