5G & Autonomous Cars: Flashy Promise Meets Complicated Reality
The latency breakthrough
Low latency is the 5G capability attracting the most attention for vehicle automation. While there's no universal agreement on what's needed, a frequently cited target is end-to-end latency of 10ms or less. Depending on how much an AV's onboard computer relies on network-based services for decision-making, the requirement could be as low as 2ms, Nokia's Beltrop said.
That's much tighter than the requirement for voice-over-LTE (VoLTE), one of the main low-latency applications of 4G. VoLTE can work with a round-trip latency of about 100ms, said Cameron Coursey, vice president and CTO of IoT at AT&T. This is where 5G becomes a windfall, because the new specification is designed to provide much lower latency, beginning with RAN latency as low as 1ms, compared with tens of milliseconds on average with 4G.
Another important piece of 5G for low latency is its flexible network architecture, which will let operators place computing resources toward the edge of the network to avoid long round trips to distant cloud data centers. More on that in a later article.
Big driving data
Other AV applications may need 5G for sheer speed. All self-driving vehicles run complex software, especially DNNs (deep neural networks), that need large updates. The cars also continuously collect huge amounts of onboard sensor data, plus information about the outcomes of driving decisions, that automakers and suppliers can collect to improve the self-driving software. These frequent downloads and uploads will benefit from the gigabit-speed wireless connections that 5G carrier networks are designed for.
Software updates may be frequent, and centralized systems may collect and analyze driving data quickly, but these are unlikely to happen in real-time. As a result, operators won't necessarily need to provide that kind of broadband to AVs in motion. Big data transfers could be activated when the vehicle is stationary and in range of a high-speed connection (at a charging station, for example), especially in areas with dense, high-frequency 5G coverage.
Another automation concept calls for AVs to share real-time sensor data so a car can "see-through" the vehicle in front of it, especially a large truck that blocks the view of traffic ahead. Because it involves real-time streaming video, this is likely to require both high-speed broadband and low latency, with assistance from edge computing, Flament said. As a result, this application would probably go over a WAN instead of a direct V2X connection, so roadside networks would need to get faster and more robust, he said.
Making sure the network comes through
Reliability is a major concern for autonomous driving applications since a loss of signal -- or of a driving assistance application -- could affect safety. One significant 5G advance to help ensure performance and availability is network slicing, which will let mobile operators set a specific quality of service for an application by assigning virtualized network resources to it. An AV application could be assigned priority over other network applications due to that safety requirement. This is one thing Ford, for example, is seeking to ensure is supported in its AVs.
If autonomous driving services need guaranteed QOS on 5G networks in the next five years, they will probably use generic network slices like those used by other applications running on shared edge servers, 5GAA's Flament said. This edge computing infrastructure, along with network slicing, will be deployed first in dense urban areas where there are enterprise customers for it. Later, when 5G comes to highways outside those areas, carriers may create specialized network slices for AVs if a road operator requests them.
Network slicing is a start, but making 5G networks reliable enough for self-driving will be a tall order, analyst Philip Marshall of Tolaga Research said. Achieving low average latency isn't enough to support mission-critical applications like driving, Marshall said. What's needed is consistent low latency. That will require much higher network density than some people expect, including redundancy in both cells and computing infrastructure -- the lower the latency needed, the higher the cost, he said. Also, large-scale implementation of network slicing could take several years and hasn't yet begun, Marshall said. To achieve results without building separate infrastructure for priority applications, it will require virtualization of both the core of the network and the RAN. To use network slicing for self-driving applications on the open road, carriers first would have to convert to a cloud RAN architecture over a broad area, and it's way too early for that, he said. "Network slicing will get incubated with applications that don't have wide-area requirements," Marshall said. "To try to implement network slicing over a wide-area 5G network environment is crazy."
New role for carriers
Network uptime will become a bigger issue when 5G networks start supporting autonomous driving, Ericsson's Herlitz said. That may start with early deployments for enterprise customers such as logistics and waste management companies.
Such customers will need service-level agreements (SLAs) better than anything offered now, particularly with regard to fixing outages, because most carrier SLAs today are geared toward consumers, he said. In this case, the mobile operator will need to be integrated into the customer's business, with technicians on site to solve problems that affect network availability immediately. When a customer's core business relies on robotic trucks, 48 hours without service is prohibitively expensive, he said.
For times when AVs travel beyond the reach of all this infrastructure, automotive supplier Continental has demonstrated a system to help the vehicle prepare. "Predictive connectivity" uses historical information about network performance and predictions of a vehicle's route to gauge where the vehicle might run out of coverage, Continental says. Then the car can change to a different network, prioritize the applications in use, or even shift to a failover mode in which it relies on built-in sensors and computing power.
As for whether mobile operators, automakers, application providers or other entities will be held responsible if a network-dependent self-driving application fails, it's too early to know, Herlitz and others said. Even without the network element, liability for traffic incidents involving AVs is already a hot topic that's far from being resolved.
Setting a timeline
Bringing together 5G and self-driving cars will involve multiple daunting technology missions for players in both networking and automotive. 5GAA is working with network vendors and automakers to converge their timelines so neither cars nor networks get stranded waiting, Flament said.
While the first 5G modems should start showing up in new vehicles in 2022, ones that support URLLC and new, high-frequency radios probably won't arrive until 2025, Flament said. Meanwhile, 5G infrastructure that can talk to those modems to enable self-driving isn't due for commercial deployment until about 2025, Flament said. A new generation of V2V that lets vehicles share more data, more reliably, to better support self-driving, may arrive around the same time, he said.
As 5G expands the role of mobile networks from primarily consumer voice and data services to emerging and mission-critical such as automated driving, network demands are growing and infrastructure becoming more complex. While the roadmaps for both 5G and self-driving are still being drawn, there's a clear possibility that 5G will help drive vehicle automation forward.
In the rest of this series, I'll look at more details about two evolving aspects of connected self-driving: edge computing and wireless spectrum.
Next page: Sidebar on C-V2X