When will Waze no longer need people? Currently, Waze users tag vehicles stopped on the side of road, debris, slowed traffic, and those pesky hidden police. However, the state of the art in computer vision, powered by deep learning, means that it should not be too long before onboard cameras will enable the vehicle themselves to sense and interpret what is happening on the road around them. (Already, users are asked to confirm what the algorithms suspect might be happening.) Today, production vehicles can parallel park or autonomously slam on the brakes in emergencies, and this is with relatively coarse cameras and simple sensors.
Imagine the capabilities when most cars are equipped with LiDAR, radar, and cameras. Already, the price is plummeting and LiDARs can now be had for $4k, a huge reduction from a couple of years ago. So, if the hardware cost has dropped — and will continue to drop — deep learning or other AI technology had advanced — and will continue to advance — and onboard computers (which seem to ignore Moore’s law and are continuing to be faster and cheaper) are more powerful, we can assume that cars will have pretty good “situational awareness” in the next 5–10 years, sufficient that vehicles can share those findings and they can be aggregated, thereby obviating the need for the Waze user to get involved. The car will be the Waze user, not the human passenger. The only concern are police vehicles covering their cars with stickers that confuse neural nets to make them think that they are toasters.
We won’t need the user’s smartphone app either. In 2021, 1/3 of cars will be connected and that is expected to rise rapidly. 5G is coming too and will offer significantly higher throughput than 4G. So, cars will have more powerful sensors and compute capabilities and be connected at high capacity. Thus, their local findings of road conditions can be sent to some centralized service, such as Waze HQ, for aggregation, processing, and maintaining real time maps.
While that processing might be fast, it might be slightly quicker to receive insights from neighboring cars — and if you are sending a warning about a danger in the road, every second can make a difference. Like the telephone game in which a message is passed from one person to another, cars could do this too. A connected vehicle that spots a blown out tire in the road, could relay its message down the road behind it, one car to another. Such vehicle to vehicle (V2V) capabilities, and more generally vehicle to infrastructure (V2I) and vehicle to anything (V2X), are in heavy development. This is because V2V, and later V2I, are key to reaping the full benefits of autonomous vehicles.
One of the major benefits of autonomous vehicles will be increased safety. If two nearby cars are communicating their positions, directions, velocities, and intentions, they should be far less likely to crash than two uncommunicative human drivers today. A staggering 1.3 million people die in road crashes each year. That’s an average of more than 3200 deaths a day. V2V can offer a significant reduction in crashes. Importantly, this capability is a “networked good.” That is, it requires a critical mass to produce benefits. As such, like the utility of the first fax machine, which OEMs or car manufacturers are going to be first and prime the (empty) network? It would add to the cost of the car but with little benefit to the driver.
The government gets this and they have concluded that they will likely have to mandate it through regulation:
NHTSA believes that V2V capability will not develop absent regulation, because there would not be any immediate safety benefits for consumers who are early adopters of V2V. V2V begins to provide safety benefits only if a significant number of vehicles in the fleet are equipped with it and if there is a means to ensure secure and reliable communication between vehicles. NHTSA believes that no single manufacturer would have the incentive to build vehicles able to ‘‘talk’’ to other vehicles, if there are no other vehicles to talk to — leading to likely market failure without the creation of a mandate to induce collective action”
What is great about this is that government may regulate minimal V2V capabilities based on safety, which is part of their mandate, but this will certainly help pave the way for wide scale autonomous vehicles. With V2V in place, can more likely coordinate and increase flow at intersections. At the extreme, they may be able to zoom through intersections without stopping, as in this kind of scary simulation from MIT:
(Rather than a fully decentralized algorithm, this and other situations may require some local controller directing traffic but with V2V in place, then it will be feasible to put such V2I infrastructure in place. My bet is that V2V comes first, then corporate V2I, and later government V2I infrastructure.)
That could increase fuel efficiency, as will the ability of cars to platoon or draft, more closely and safely than human drivers can do today. Again, this can increase fuel efficiency. With a set of high-situational-awareness vehicles roaming the streets, internet connected, crime may also be reduced.
So when is this happening? Autonomous vehicles are definitely coming; not tomorrow, but they are coming. While Google, Uber, and Tesla get all the press, the reality is that dozens of companies are getting into the fray — from the major car manufacturers (GM, Volvo, Daimler, Nissan, Audi, Toyota, Ford, PSA etc), OEMs (ZF, Delphi, Denso, Continental, etc), to internet companies (Baidu, Google). They are spending significant amounts of R&D cash and placing big bets.
Most estimates place majority AV roads at around 2050. The transition from horse to car took 30 years, one piece claims as justification. At some level, it is far easier to imagine the fully AV world — cars zipping around, racing through intersections without stopping — than it is to imagine the coming mixed and likely messy transition phase over the next three decades.
Source: Deep Learning on Medium