Hacker News Reader

Comment by bradleyjg | Hacker News Reader

The request for preliminary injunction, available here: https://www.documentcloud.org/documents/3515651-Waymo-Motion... only has to do with LiDAR.

I'm far from an expert on the self driving car race, but I think even if Uber had to abandon LiDAR altogether it wouldn't mean it would be impossible to approach the problem in a different way. And it isn't clear that if the injunction were granted they'd have to abandon LiDAR altogether -- it may be that they could try a slightly different approach and/or develop a second firewalled team from the one tainted by Lewandowski.

It would surely be painful but I don't see why it would mean shutting down the self driving car project.

View on HN

The problem isn't the technical side of things, it's the legal side.

If Uber loses this they probably won't be able to hire from Google, they will have to do "clean room" implementations, and this will all have to be done on the up and up with documentation and proof that whatever product they come up with was entirely started after the ruling with no use of anything that could have come from Google or Waymo. Not only will that all be a huge expense, it will be practically impossible to compete with those barriers constantly looming over the R&D team, not to mention all the extra legal red tape.

That is the crushing blow, not that they can't use LiDAR

Also, famously, Tesla does not use LiDAR. Not to mention, there are plenty of 3rd party LiDAR vendors whose equipment could be used to substitute the design in question.

If you think about it, the only sensor you need is a camera. Humans control cars with just their eyes as sensors. Our brains just possess superior perception. Maybe if uber pushed the boundaries of computer vision, they could make progress toward vehicles with simpler sensor suites.

"Yeah, so this is using only cameras and GPS. So there's no LIDAR or radar being used here. This is just using passive optical, which is essentially what a person uses. The whole road system is meant to be navigated with passive optical, or cameras, and so once you solve cameras or vision, then autonomy is solved. If you don't solve vision, it's not solved. So that's why our focus is so heavily on having a vision neural net that's very effective for road conditions." - Elon Musk, latest TED (https://www.ted.com/talks/elon_musk_the_future_we_re_buildin...)

I'd love a really well implemented augmented reality windshield though, one that added sensor data to my field of view.

Imagine being able to see heat signatures at night, or through obstructions.

Not really a strong argument. Birds prove that you can fly with only deformable surfaces (flapping wings), and in particular without dedicated devices to inhale air and push it out fast (engines). But if you had to build airplanes with deformable surfaces but not engines, you would never be competitive (at least for first century of aviation technology...).

Which proves -- in the case of flight -- that an "unnatural" refactor can be better along certain dimensions than what nature came up with: airplanes are better in terms of average speed and payload transported by air per unit energy consumed.

The parent's point was that, for driving, the "natural" human way is still better along one relevant dimension [1]: how much sensory data is required (SDCs need a lot more).

SDC engineers would regard it as an improvement if they could get SDCs to work using only the sensory data that a human would get. The parent was saying that Uber could show superiority by similarly doing more with less.

[1] Humans also drive with coarser control over the car i.e. no direct computer interface, just gripping the wheel and pushing the pedals.

The parent is clearly saying that this is a plausible route for Uber engineers using the biological proof-of-concept, which I have rebutted. They are not just saying it would be an improvement.

You cannot rebuke that claim via an aerospace analogy because they are not equivalent fields. It remains to be seen whether lidar is better for cars than "vision".

"Undermine" would have been a better choice of words than "rebut". Indeed, my first comment claimed explicitly to show his wasn't a strong argument, and I think that is obviously the case even if the fields aren't identical. (When are they ever?)

Bio creatures can also fly at low speeds, powered by bugs. Planes get nowhere close.

Passive sensing lets your eyes see stars and comets: long range, power efficient, avoids sensor conflicts with other objects.

"Our brains just possess superior perception."

This is the kicker. Lidar allows companies to make up for a lot of the current shortcomings of state-of-the-art computer vision.

As to how easily Uber could advance state-of-the-art computer vision, CV is an academic field with decades of research efforts built up already. To be fair, CNNs are still young-ish in terms of producing results, but Google, FB, etc all have better teams than Uber in this area. The fact that every major self-driving car project except Tesla is using lidar says a lot about how insiders see the first-to-market strategy playing out. Relying on pure CV at this point is kneecapping yourself.

Right, all you need is human level AI, seems simple enough.

We humans also use acoustic input as well as tactile input, acceleration sensors, etc. It's certainly possible to use vision only, but pretending that humans don't use any other senses when driving is simplifying things a bit.

It would mean losing years of progress and wasting large amounts of money that they spent on now-useless research.

Reply on news.ycombinator.com