I actually think that low speed driving and parking in very narrow streets is much easier for an autonomous system than for a person. All those obstacles are made of stone so nice juicy returns for the sensor systems. Carrying out a turn with a centimetre of clearance from a hard surface is much easier for a computer than maintaining lane discipline on a road with faded markings at high speed.

A lot of impact avoidance with other vehicles and pedestrians is done by communicating with looking at each other though. The easiest solution, car always yields, would just mean it will take forever to move because people will lose all respect and just treat it like a solid obstacle. Also, sometimes cars can and do pass within a few cms to humans (ok if everyone is aware of the maneuver and it's slow), would an automated system do that?

> The easiest solution, car always yields, would just mean it will take forever to move because people will lose all respect and just treat it like a solid obstacle.

I've often wondered how self-driving cars will deal with this. Imagine we reach the point where self driving cars are perceived as being guaranteed to stop before hitting a pedestrian.

This presents a lot of problems that need to be overcome. A car-jacker could rob you simply by getting an accomplice to stand in front of your car. Teenagers might start jumping in front of cars as a high-tech game of chicken. What happens when somebody steps into the road and the car brakes so hard it gives the occupants whiplash?

I'm sure all of these problems will be solved in time, but it's going to be interesting to see how!

There is a 500 m long street in front of my house where it is customary that people cross it everywhere and cars are just expected to stop for the people. There are no traffic lights or pedestrian crossing or any special rules, its just that people started doing this decades ago and now it is just tradition to do this.

So now you have a long street where tons of people are crossing every 10 meters during the day because it's a quite busy shopping district.

Although cars are driving slow, it must be very hard for autonomous cars because there are people everywhere running in every direction.

Don't forget computers have vastly faster response times than people. A car going 5mph with 10ms reaction times can stop in ~1 foot. http://www.random-science-tools.com/physics/stopping-distanc...

Do you want to be inside the car when it does that? It is not going to be much fun unless you are paying attention and bracing yourself, in which case I'm not sure I see the advantage of letting the computer do it for you.

You are probably overestimating the jerk from that. 5mph has (1/12)^2 = 1/144th the energy of 60mph. Passengers often don't pay attention to what's going on and breaking force is generally not enough to really mess with people.

That said, the default is not going to be stopping within 1-2 inches of pedestrians, rather it's going to gradually slow down to deal with pedestrians while trying to maintain forward progress. The point is a self driving car could be more aggressive than most people while staying safe instead of freaking out and just stopping because people are jaywalking.

You're right of course. It's hard to imagine any useful vehicle spending much of its time moving at only 5 mph, though, so I suppose I'm really imagining what such a stop would be like from a more normal travel speed like 20-25 mph.

I am picturing this more as an extreme edge case than anything else. If a person can safely do 20-25mph in an area then a self-driving car should be able to do the same thing.

It's the edge case where a car needs to cross a really high traffic area that's going to be an issue and people can't do 20mph right in front of a high traffic box store entrance etc. Further, going slow for a few hundred feet really does not take that long.

I would VASTLY prefer to be in that car than to be in a different car with a human driver and watch a person get killed by the vehicle I'm riding in.

Nobody is going to die from a 5 mph collision. It's not even likely to break any bones. At 30 mph, the chance of death is still only 5%!

There may still be reasons to prefer robot cars for low-speed environments but the probability of such a vehicle striking and killing a pedestrian is already very low, so that is not relevant to a discussion of a robot car's ability to stop very rapidly at very low speeds.

http://www.nhtsa.gov/people/injury/research/pub/HS809012.htm...

The "every direction" bit is actually something autonomous cars handle far better than humans since they can have eyes pointing in every direction. The hard part is noticing that someone waiting at a crosswalk wants you to stop so they can cross.

The problem is not necessarily "obstacles made of stone." Think about lanes in the countryside where you only see grass and the possibility of a ditch to either side... I guess that is really complicated for a self-driving car to manage in each and every case as well as human does.

Yes, walls are easily located. Holes not so much. Emptiness is quite difficult to see.

Except that it's much more likely for there to be pedestrians.

Why? What specifically is hard about narrow streets? That actually sounds like the kind of thing a self driving car could excel at.

The bigger problem, is imagine, would be weather. I dunno how completely they've solved that problem yet.

I recently drove in freezing storm. Whole frontend was covered in icy-slushy mixture. Parking sensors were freaking out and working totally randomly in such conditions.

If I ever get an autonomous car, manual driving is a must-have feature :|

I once had a car that didn't work at all when it was below about 32 or above about 70.

Keeping my horse was a must for those situations.

I once rode in a plane that could only go a few hundred miles before refueling.

Keeping my car was a must for those situations.

I once rode in a train that could only go ~60 mph.

Keep my car was a must for those situations.

(hint, I'm talking about early stuff)

One of many examples would be a single lane street which goes up a hill and around the corner. Although there is only space for one car you can drive both directions. Rather often than not you have two cars meeting in the middle and then one car has to drive back to let the other car pass. This is pretty difficult to do as a human and I am not sure how an autonomous car would do that.

The same way it goes forward?

2 autonomous cars could negotiate which has the shorter trip, an autonomous car encountering a human could default to timidity.

Oftentimes there's a two-way street when only one car can pass. Sometimes there are signs for that, but not always. And sometimes you need to check mirror to see if car is approaching from behind the corner before you can proceed.

When a human messes up its a cheap accident, just bad human judgment call number one billion.

When a self driving car messes up it'll be a very expensive lawsuit for pain and suffering and psychological trauma of not being in control while the accident unfolds and surely its all the mfgrs fault for having a software bug.

Its not so much the physical characteristics of being narrow as being a location where many accidents happen compared to an interstate-class road.

I doubt that'll be anything like what the outcome of an accident involving a self driving car will be like.

How will they negotiate a stalemate? Against a human or a non-human opponent? Humans will use headlights, gestures, eye contact.

One scenario that I keep thinking about all the time: traffic in most south-east asian countries like Vietnam (Hanoi for example). The streets are occupied by a mix of cars, pedestrians (+ animals) and motorcycles. Everything moves ultra close.

Must be a nightmare for every car proximity-sensor.

Actually, no. Speeds are low enough that almost-zero reaction time is a huge advantage, and the rules are basic swarm-like behavior (plus bigger cars have priority).

I'm not sure. In my experience, you have traffic situations where everyone around you is trying to weave in and out through the smallest gaps. The car will need to get really good at predicting behavior or drive so conservatively that all other vehicles end up passing you by.

Lol, yeah. Asking for a self driving car programmed to follow traffic laws that can handle Saigon or Bangkok is like asking for a smart drone that can fly safely thru a tornado.

Is it? Can an experienced airplane pilot even fly safely through a tornado?

I agree that this is a really bad example. Especially as I was in Hanoi 5 days ago.

I witnessed 2 car accidents when I was there and both were vehicles turning (slowly) and clipping motorbikes in their blind spots. Not gonna happen with a computer.

    > and it is already difficult to navigate those places
    > as a human

I don't know why a machine that has 3d perfect vision and knows exactly how big it is relative to its surroundings would be worse at this that a human bean.

Because in a lot of situations you are not alone on the road and you have to negotiate with other cars or people and decide what to do. There are a lot of single lane streets which can be passed in both directions and the car has sonehow have to deal with the incoming var on who should drive first or who drives backwards.

Meh, again, I don't know why software that's never having a bad day, never frustrated, is trained to drive defensively, and has the actual highway code programmed in to it will struggle more than Pedro will, especially given that if the other car is also autonomous then they can actually negotiate.

People always raise these issues... "I know about a tricky street that no computer could drive!" But it doesn't really matter in terms of selling autonomous cars. If autonomous driving is only enabled for the top 90% of the most autonomy-friendly routes, then only people in those areas will buy the first round of autonomous cars. Over time, AI improves, maps improve, and people put pressure on municipalities to improve roads.

But is any of that a problem for Tesla to sell a lot of autonomous cars? No. Not even a little.