Comment by toomuchtodo | Hacker News Reader

Comment by toomuchtodo | Hacker News Reader

Doesn't appear to differ from previous Tesla autopilot accidents where the vehicle couldn't detect a tractor trailer crossing the path of travel with cameras or front facing radar. Interesting that Autopilot had only been activated for 10 seconds prior to the incident occurring.

I am not a domain expert. Front facing radar and cameras alone is clearly inadequate for forward facing object detection [1] (unless we're going to mandate radar detectable skirts and images superimposed on the sides of tractor trailers for vehicle detection). Would solid state LIDAR (essentially laser ranging I suppose, as you're not building a 360 degree point cloud) solve this particular edge case? Some automakers are including laser headlights in their vehicles [2] (although I don't believe this is yet approved for the US market); would it not make sense to convert the headlights into front laser ranging sensing systems along with their illumination function?

[1] https://www.caranddriver.com/features/a24511826/safety-featu...

[2] https://www.osram.com/am/specials/trends-in-automotive-light...

I also found the 10 second engagement time odd. On the one hand, you might not thing it's enough time to become seriously distracted. (eg, you wouldn't be "comfortable" with the semi-autonomous operation yet).

On the other hand, the engagement of autopilot might have been caused by a desire to explicitly NOT pay attention, however briefly. think: "ok, now let me read this text from the person I'm on my way to meet."

I have found myself occasionally using autopilot to augment my driving when I'm doing something that inherently decreases focus on the road, however briefly; eg, taking my sunglasses out of the glovebox. I still drive the same way I normally would in any other car when reaching for something in the glovebox, so it's easy to think "any enhanced safety is better than none".

On the other hand, it could lead to a false sense of security where you pay 10% less attention than doing the same task in a non-augmented car. Or you get comfortable doing such things, so you do them more frequently than truly necessary. (for example, you might just shield your eyes from the sun with your hand rather than grabbing sunglasses in a non-augmented car).

Tesla already has RADAR. Why would they need LIDAR?

The issue isn't with the sensors, but rather the system reacting to sensor data.

Radar is not sufficient for mitigating this scenario. Too low resolution.

You can buy much, much nicer radar than Tesla specs for their vehicles.

The stuff the DoD buys (or was buying a decade ago) is good enough to tell two basically identical tractor trailers apart based on minor features, dents/dings, etc.

Then why do automakers who support AEB not spec this radar? The Car and Driver link in my GP post above this specifically demonstrates that Tesla is not immune to failures to detect obstacles in the vehicle's path; it happens across many automakers (Subaru, Toyota, etc).

Probably because it hasn't come down in cost yet and/or there's physical limitations involving packaging the receiver.

As the long slow march of progress continues it may come down in cost enough to be viable on vehicles.

Can you share some more links to the radar technology you mention with high resolution and discrimination functionality? I am interested in performing more research on this topic. I'm familiar with several phased array radar platforms, so I might be aware of this specific implementation, but perhaps not!

https://www.baesystems.com/en-us/product/vehicle-and-dismoun...

As I understand it it's the algorithms that analyze the radar data that are the secret (literally) sauce. I didn't work on any of that. I just plumbed various data from A to B. I don't know much other than it was really damned good for being "legacy" at the time I was working there.

Were the tractors moving at different speeds in this test?

The test data was recorded during an average day over a medium sized city in the DC ares.

So yes.

I don't think we can conclude that cameras are inadequate for forward object detection. Only that the processing systems used so far were. This can be a matter of training of the neural networks and the amount of processing power used in the computing units. Radar is limited, as long as trailers in the US are allowed to have such high gaps to the ground. A lower bar on the sides would both help the collision detection systems as well as reduce the crash itself.

They could probably solve the problem by mounting the radar on a trunion so it can sweep up and down (over say ~30 degrees).

>A lower bar on the sides would both help the collision detection systems as well as reduce the crash itself.

The Neural Network should be capable to recognize a truck from any angle and even if parts are obstructed. I hope we will get the details public and see if it actually tagged the obstacle but ignored it or it detect it as something else like maybe a side road advertising panel

> "Radar is limited, as long as trailers in the US are allowed to have such high gaps to the ground."

One of these crashes involved a firetruck, which unlike a semi-trailer, does not have a large gap underneath it.

Radars are limited period. The radars used by Tesla lack the angular resolution to distinguish between a stationary object next to the road (e.g. a building or parked car) with a stationary object parked right in the middle of the road.

But Elon told me I will be able to rent my Tesla out as an autonomous cab by the end of this year or so.

The radar Tesla spec'd is quite limited.

The radar that companies people on HN don't think highly of sell to the government is "scary good" and that's just the decade+ old tech that I'm familiar with.

I'm sure there's reasons Tesla doesn't/can't put that in cars though.

It's nice and all that the military has such cool radars, but how much do those systems cost? How large are they? What sort of power requirements do they have? What sort of duty cycles do they have? And can they legally be exported?

Elon caving and adopting Lidar is probably more likely.

One of these crashes involved a firetruck, which unlike a semi-trailer, does not have a large gap underneath it.

That fire truck wasn't crossing the road, it was at the road side, which creates a different situation. The problem with high trailers is, that they very much look like a bridge to a radar. In any case, a proper side bareer can prevent deadly accidents.

It doesn't look that way, seems to have been in a lane, but any report on whether the autopilot was in fact engaged?

https://twitter.com/CC_Firefighters/status/95552999131956019...

If the radar cannot distinguish between a bridge with five meters of clearance and a "bridge"(trailer) with a single meter of clearance, then the radar simply lacks the angular resolution to be effective. It's the same exact issue as it being unable to distinguish objects on the side of the road and objects in the middle of the road.

It's insufficient angular resolution for the application, no matter which way you slice it.

Reply on news.ycombinator.com