Comment by Animats | Hacker News Reader

Comment by Animats | Hacker News Reader

There's heavy political hype around "5G".[1] There are claims that it will matter for self-driving cars (despite Waymo saying they don't need more bandwidth), medical applications, and "smart energy". All of those are bullshit.

The "if only we had more detailed maps, self-driving would work" is a fake argument. Once you have enough info to know where the fixed trouble spots are, more data will not help you. Trouble comes from unexpected moving objects.

(2004 (not 2005) DARPA Grand Challenge, where everyone did badly. CMU team tried to do it by manual pre-planning. DARPA gave out the course as a set of waypoints on a CD 2 hours before the start. CMU had a big trailer full of people at workstations to plan out the exact path in those two hours, using high resolution aerial photos. But the USMC Colonel in charge of the event foiled their scheme. Just before the event, a few of his Marines went out in the dark and put up some obstacles the vehicle would have to go around. And, sure enough, the CMU vehicle plowed right into a sheet metal fence and got stuck.[2])

What we're likely to see is bandwidth that changes drastically as you move. You'll get great bandwidth with line of sight to a nearby base station, and then it will drop off drastically as you get further away. That's inherent in using 26GHz for high bandwidth.

The likely benefit is that it becomes possible to provide enough short-range bandwidth for many people to get video-rate bandwidth in crowded areas. So people can watch the game on their phone while in the stadium.

[1] https://thehill.com/blogs/congress-blog/technology/420509-wh...

[2] http://overbot.com/grandchallenge/images/race2004/cmucrash.j...

The worse thing is, if you're in the middle of nowhere with 0 reception what then? We need to rely on what's available locally above anything external, of course downloading all the map info and caching it definitely should be a big priority. I wonder if based on some location information and given that the car can 'see' where it's driving if it will improve GPS clients and work even when reception is so poor. I hate driving and hearing "GPS connection lost" from Google Maps, like really? You can sense certain things about my phone, why can't you just keep going, I've pre-cached a good chunk of the map too!

I'm not a GPS expert by any means, but dang.

Then the government will add new legislation that you are not allowed to go anywhere where there is 0 reception :)

GPS doesn't require a data connection.

But it does require a connection to GPS satellites.

GPS is receive only. So there is no "connection" per se, "reception" might be better here.

And you’ll have reception anywhere on the planet as long as you’re outdoors and not in a tunnel or a deep canyon or similar.

Urban canyon was a fun term from a previous phone-support gig...

The bandwidth isn't just for downloading maps to cars.

It's also to enable autonomous vehicles to constantly stream data and video feeds back to humans who can take over in real time if the car doesn't know what to do.

That's completely unrealistic and unsafe. Even with widespread 5G deployment the cellular network will still lack the reliability necessary to make real-time remote driving feasible. When happens when the network drops packets during a critical maneuver?

Furthermore you can't expect a remote operator to suddenly take over with no context of preceding events and immediately control the vehicle in a safe manner. It takes a little time for anyone to understand what's actually going on.

It could be good enough if it’s a backup system only. Then again, probably the only reason the government would be enthusiastic is because these amounts of bandwidth would enable continuously streaming GPS tracked dashcam surveillance from every citizens car.

It's not nearly good enough for a backup system. The whole point of having a backup system is for safety and reliability. You can't build a viable backup system atop an unreliable network.

This plays into the density of towers. 5G has a very limited range unlike 4G. It's going to take years (if not decades) to get to a place where 5G can cover the entirety of the US.

Even if we had 100% 5G coverage it still wouldn't be adequate for a safety critical system. What happens when the tower fails on a hot day because the cooling system broke, or a construction crew cuts the backhaul fiber, or the carrier has an infrastructure failure because someone spoofed their BGP routes?

If level 4+ autonomous vehicles are ever going to work then at a minimum they need to be able to operate safely with zero network connectivity.

Imo they need to BE the network

Dynamic mesh networks can be a nice supplement but no one has ever demonstrated large-scale reliable operation with mobile devices. And you can't always count on having another vehicle in radio range. So turning vehicles into the network won't make remote control viable.

I think the idea would be to have the vehicle detect that its out of its operating spec, and park until either

a) situation returns to within operating spec b) remote control comes through

The remote control aspect would be severely limited and intended only to get the vehicle back to a situation within its operating spec. This trades the very difficult problems 'need to immediately control the vehicle in a safe manner' and 'what happens if a remote control is lost' into the difficult problems 'how do I detect I'm out of operating spec' and 'how do I park safely when I'm out of spec, especially if I've just lost remote control'

Park where?

I frequently drive on rural roads where there are blind curves and no real shoulder (and not much cellular coverage either). A human driver can usually steer a disabled car mostly off the road by going down into a ditch or up an embankment. But it's still not a safe place to be. And autonomous vehicle software is nowhere near being able to handle those maneuvers.

I mean yeah, thats part of the challenge. Its still an easier one to solve than the one you posited

How is it easier to solve? It seems about as difficult to me, assuming the requirement is to maintain a high level of safety in all situations.

This already is happening with several autonomous vehicle companies.

It helps with municipal adoption.

Is this comment serious? How could that ever work??

This is actually Waymo's strategy that Waymo is allegedly actually using in Arizona, so presumably it is serious. (Though I agree it's a terrible idea.)

I want to be able to drive a car from the safety of my home using a VR headset. They can turn actual trucking into a trucking "simulator".

Oh I thought the main argument why 5G helps autonomous cars was the lower latency.

That is a horrendous idea. 5G isn't going to be some magical technology that's always on all the time. It will have outages and black spots just like any other.

> It's also to enable autonomous vehicles to constantly stream data and video feeds back to humans who can take over in real time if the car doesn't know what to do

If this is required for autonomous vehicles, then I'll pass.

That's the reason very few companies are talking about selling those, and the entire discussion is about cars as a service.

Given enough time, they will probably stop requiring external help, but it's very unlikely that the first autonomous cars to run without it.

I often see latency being touted as the game changer for 5G. With compute being pushed out closer to customers "Mobile Edge Compute"

Could anyone enlighten me to what the use cases are? Because it's still dubious.

> With compute being pushed out closer to customers "Mobile Edge Compute"

That seems implausible to me. We already have the equivalent of supercomputers in our pockets, and yet the trend is to move computing away from the end user, onto centralized servers.

In some ways but not in others. Edge Computing is actually a pretty hot area in telco but also other areas. One of the issues is that there’s so much data being collected, some of which needs real-time action, that you need distributed computing to filter and take actions based on rules. This edge computing can be fairly substantial.

The most instantly recognizable one, with a clear commercial potential, is cloud gaming.

Other fuzzier ones are, uh, crunch live video or audio captured by the device via beefy compute to... display overlays? Search similar products? Live translation? Calculate evasive maneuvers? A lot of these are limited by interface concerns that introduce human or other 'latency' anyway, making connectivity latency between the base station where the 'edge' is, and the device itself less of an issue.

The latency needs for online gaming aren't as stringent as you might think.

I have a Samsung TV which, if it is not in "game mode" introduces much more latency than my DSL internet.

I couldn't win at Titanfall at all with the TV in normal mode, in game mode I was able to progress halfway up the leaderboard.

So far as I can tell, "Edge" is for people who think AWS isn't expensive enough and they'd rather pay AT&T or American Tower prices.

Except I have 4-8ms ping on my internet, with rare peaks at 20ms.

I have between 40ms and 200ms on 4g. Average around 80, with often peaks above 300ms.

With the gaming scenario some distinction between display latency and network latency is worth reviewing. Network latency is something that software can cover for in some degree(to create online experiences there's always some kind of synchronization of different wallclock timelines going on). Display latency is universal, though. A faster display will let the game give you feedback faster, so long as we are talking about feedback in the local experience(so: aiming, movement are commonly predicted on the client, while hit registration often uses a server roundtrip). This separation of responsibilities allowed people to play Counter-Strike 1.x on 200ms dial-up connections back in the 2000's.

When it's a pure video feed you get the impact of all the latency all the time, and that means the problem has to be solved with brute force reduction of bottlenecks everywhere.

So how much latency is enough for responsiveness? This is pretty easy to derive from the common target framerates: for 30hz display, you need 33.3ms. For 60hz, 16.6ms, for 120hz, 8.3ms. Since perceptual latency is known to keep improving up through 144hz it's reasonable to say that we should be looking for only 5-6ms at most, while most broadband connections are still achieving pings in the 20-100ms range, depending on the game and the specific connection. On the very best connections, that is, we can assume a "30hz transparent" video stream, which is fine for casual experiences but severely impacts competitive ability when tested. In many current popular titles this manifests in mechanics that are both latency sensitive and require a server roundtrip, e.g. building in Fortnite.

Yeah I'm totally going to tell you what I'm currently working on because I like random competition so much :)

really though, the waves aren't faster than lightspeed. Although lightspeed in different media might depend on frequency, indeed, I do not think that's the problem, rather than congestion, because the refraction index of Air is close to 1.

Congestion is a typical problem, though. ""Mobile Edge Compute"", what about "capsule networks"?

It just means that the last mile over the air will only add 1ms of latency. Compared to LTE which I think has around 5ms. The rest of the network (read Internet) is still just as fast as before.

If you think of data that's one thing, but Radar a rather different one. You won't need to bump a scanning ray around, if you can simply detect changes in the static urban microwave background. You'd need many senders. Suppose those fit into lanterns, that have line of sight by definition. I'm fantasizing about MEDs emitting 26 GHz with 1GHz duty cycle, lol.

In my view, 5G primarily benefits carriers. I'm not personally excited about it, as it won't provide anything I need that 4G doesn't currently provide. I won't be going out of my way to upgrade anything with 5G in mind (but I won't be avoiding it, either).

I never heard about more detailed maps as a requirement for self driving

Perhaps those people should check how the first navigation device worked ( FYI: without GPS back in 1981) as inspiration to see if there are other ways.

https://jalopnik.com/the-first-commercially-available-car-na...

Why did DARPA give out the course ? Was that part of the challenge or a way to favor a contestant ?

The 2005 DARPA grand challenge was "follow these dirt roads through a desert" where they wanted the vehicles to follow a specific set of dirt roads.

The self-driving car problem can be divided into several chunks, and the 2005 Grand Challenge intentionally tested some of those chunks and ignored others - a divide-and-conquer approach to problem-solving any developer will be familiar with.

The subsequent 2007 urban challenge added extra chunks, like other moving vehicles, high-level route planning and replanning around blocked roads, manoeuvres like parking and three-point turns, and suchlike.

Broadcast has other solutions like broadcast. It's when everyone has different things you need 5G.

Reply on news.ycombinator.com