Comment by jameskilton | Hacker News Reader

Comment by jameskilton | Hacker News Reader

This has been true for a long time. Radar can't differentiate between a wall, a stopped car, or a dead skunk. Heck, it can get confused with a bridge's shadow on a bright day. Thus autopilot has to be configured to deal with these situations: stop or run into them. The vast majority of the situation you don't want the car to stop, as that will lead to a more dangerous situation and increase the chance of a collision.

It's one of the many situations that show that "autopilot", particularly Level 4 and Level 5, have a very, very long way to go. Also, as the reddit thread discusses the psychology of autopilot, this is why I don't want self-driving cars yet. Humans are horrendous at paying attention when they aren't directly in control. Airlines have been dealing with this problem for decades. Tesla has even started calling their services "driver assist" instead of autopilot to help combat this problem.

And yes, many people use autopilot without a problem every day, but you have to remember that it only takes one single miss to kill you (see: driver killed by ramming into a lane divider on autopilot).

> Radar can't differentiate between a wall, a stopped car, or a dead skunk

"Radar" absolutely can.

What can't differentiate between these things is the radar in Teslas that was intended for use in adaptive cruise control. There are any number of not insanely expensive radars that could do this job perfectly fine. Tesla just decided to go "full self driving" with this rather bad radar because "humans can do it with vision only!!!1!!!1!"

I don't the specifics of Tesla's radar implementation but what I have seen of the accidents usually seem like a textbook examples of incorrectly implemented clutter filtering. It's not an easy problem to solve in practice (re: cheaply), but known solutions have existed for decades.

The radar module in Teslas uses the same filtering strategy as most adaptive cruise control radar modules, which is "if relative velocity to object is nearly the opposite of the car's velocity, throw it out."

It works pretty much as designed.

Right, sounds like they drop a bandpass filter on the down-converted Doppler spectrum. It's the simplest way to "solve" the clutter problem by just getting rid of everything stationary.

Automotive radar is clearly not up to the task of differentiating stopped cars from signs at a range far enough away to brake safely, it's true. In theory if you could fit a narrow beam phased array radar on a car it could do the job but that isn't something at all in the cards. But self driving cars with lidar have enough precision to deal with this problem even if they do cost more.

Rather than showing a problem with level 4 or level 5 automony I'd argue that this, and the psychology of driving, show that level 2 and level 3 are too dangerous since they require constant driver attention without constant driver engagement. We should just abandon the idea of having those autonomy level for cars and go straight to 4 and 5. That's what Waymo and other responsible self-driving development teams are doing and it looks like it'll work.

A better approach would be to focus on making more and more improvements to level 1 autonomy. The driver would be hands-on all the time, but the computer would intervene as necessary to prevent collisions. We already have some of that with forward collision prevention systems, blind spot warning, stability control, etc. Just keeping improving those systems by adding more sensors and authority over more vehicle controls.

Level 4+ autonomy is only going to be viable on a few limited roads for decades to come. The AI and sensor challenges are tremendous. We shouldn't let that long-term goal distract us from real safety improvements that could be realized relatively quickly with minimal technical risk.

Can someone with more computer vision knowledge explain why we can't teach a machine learning algorithm to detect stationary cars.

I have seen other Tesla demo videos where it clearly detects stationary cars. It seems ridiculous to say that it is a shortcoming of camera when our own two eyes do it.

It's not that it either always detects it or never detects it with video so much as that it doesn't detect it nearly as frequently as a human would. By contrast radar is pretty much always going to see things that aren't stationary.

But it would seem to me like that is just a matter of needing more training, or maybe more processing power.

At some point with enough training data it will detect a stopped car and react much more quickly than a human would. And even if it's not perfect, all it has to do is beat a human to be worthwhile.

It's not so easy, because you don't just have to detect stopped cars, you have to reject things which aren't stopped cars. False positive rejection appears to be something both Tesla and Uber are struggling with. (In the Arizona fatal accident, the Uber vehicle had Uber's own emergency braking system disabled because it was throwing up too many false positives.)

If there was a drug which reliably blocked memory consolidation, like a more polite and controlled benzo blackout, maybe humans could act as their own self driving cars

Reply on news.ycombinator.com