Hacker News Reader

Comment by lolc | Hacker News Reader

"Beginning in early 2018, we’ll start rolling out a door-to-door self-driving taxi service to residents."

Now? Isn't that a bit optimistic? To me it looks like they have just one car for testing. At least they understand that they need the environment mapped to a very fine degree.

View on HN

We are an optimistic bunch, and it will be a slow rollout (constrained by vehicle build-outs), but we also have experience launching such self-driving taxi services.

We have a service that's live today in San Jose that serves tons of (repeat) riders daily. We've learned a lot there that we're applying directly to Florida.

How autonomous is the service in San Jose then? Do you mean to say that you have cars without drivers that serve ride after ride and come home to charge? If so I'm far behind the curve.

We still have Test Drivers behind the wheel as a last resort, but we regularly give passenger rides with zero or low disengagements.

As a real-world example: we have a couple who uses our self-driving taxi service multiple times a day every day (The Grims), and their rides are ~95+% autonomous.

You mean 95% of their rides are entirely autonomous or their typical ride is 95% autonomous?

I remember how voice to text got to 95% in the nineties. It's used for a lot of things these days, still not for safety-critical tasks.

A typical ride is 95% autonomous.

A long way to go for sure, but having been at this just less than a year, we're feeling good!

So how do you get rid of the human backup driver? Until they go, this technology has negative cost-effectiveness. Remote manual control at slow speed for difficult situations? Dispatching a backup driver?

Relying on fine-grained maps always seems to me like the Wrong Way to do these things. Fine-grained maps are guaranteed to be wrong, so you can't rely on them. And if you don't rely on them then you don't need them.

It's how you start. At some point computers will have general cognition at a level that enables them to understand typical traffic situations and improvise. Right now the road network is not standardized enough for them to rely on their sensors. One way streets? Need a plan. Turn regulations? Plan! Interpretation of lights? Hah, plan. Go over a curb? Not in the plan!

They simply can't take those decisions. Yet. As companies such as this make headway, roads will become more standardized to help them along.

It depends on how fine-grained we're talking. If you rely on your map for absolute perception (every object must exist in your map in its current real-world state), then yes, you're in trouble. If you rely on the map as a basic prior, and perceive the world through different means, then a map (to me) is the right way to go. You avoid a whole category of computer vision in doing so.

The good news is that a point cloud map is relatively trivial to create, and the pipelines to update such a map over time are becoming more understood. There's lots of startups taking on this challenge.

Reply on news.ycombinator.com