Hacker News Reader

Comment by cromwellian | Hacker News Reader

ML without hardware side radars or redundancies like high resolution maps in the areas they operate is what isn’t going to work.

Name a deep neural net with anywhere close to 100% object recognition precision/recall.

You’re engaging ins hand waved reality distortion from musk that somehow their cheap low quality sensorsuite with gaps that they shipped hundreds of thousands of, can be patched over by software and be made to drive across country and service millions of people in robotaxis one year from now despite the fact that they haven’t even filed any significant l4 miles with disengagement reports?

This seems like a market rush disaster waiting to happen. You don’t ship mvp/beta of software that can kill people so you can gather training data.

View on HN

1. You can get 89% accuracy with a convolutional LSTM https://arxiv.org/pdf/1411.4389.pdf

2. You get multiple bites of the apple the chip is processing 200 images a second. It had plenty of sign in the next frame to identify an issue.

3. All Tesla have redundant side sensors and radars, and an adversarial system for checking it's conclusions.

4. Having don't actually need to file those things in Tesla case. Why would you? They could easily test the functionality and tell no one and not worry about the regulation until they've passed a certain amount of accuracy.

5. The beta is navigation on autopilot and it's proven to be safer than regular drivers by far.

And most importantly what do you think will happen if elon release. Self driving on local roads in a year. Other than that what more would the car need to do. Be able to do that without people touching the wheel. If he just release that as a "feature" and it works there won't be a marketing issue. In fact he already has.

1. 89% != 100%. If you miss a pedestrian and hit them, you're in trouble. Do you really want a car that misidentifies things 11% of the time? The kind of nines you're looking for in a self driving car are much higher.

2. Multiple frames misinterpreting something the same way doesn't help you. Also many times you can only draw 1 conclusion from many frames e.g you need N seconds to tell a person's intent.

3. Redundant sensors are great. Radar + camera can help give good estimation about where things are in the world, but cameras have disadvantages (e.g occlusion, low sun angle, etc) and radar too (resolution, poor range, interference, etc). If you don't have redundancies for the weakness of each sensor, you are going to have problems in the long tail.

4. Ghost mode driving doesn't actually capture the long tail of crazy situations, especially since their current sensors most certainly don't store all the data needed to recreate the world. You need to drive a verrryyy long time to see some real-world situations that break your assumptions about your sensor robustness.

5. Highway autopilot is more or less solved (with varying safety bars). Surface streets are the hard part and Tesla hasn't demonstrated any progress there beyond PR / hype.

Human's aren't 100% either

1. Humans aren't 100% They aren't even 89%. They are only 76% accurate. https://www.entrepreneur.com/article/283990 and they do just fine.

2. Multi frames means you don't misinterpret the same way. The rotation scale, and other variance change.

3. You do have redunancy in the sensors and something that no other sensor besides a brain has. The ability to predict items directions and speed relative to yours. Something no amount of sensors gives and which is the real goal.

4. You should point that argument to waymo not tesla. They are the only ones looking at all the cases they can and learning from that data.

5. Highway piloting is solved in tesla case. You know who hasn't solved it no other automaker. No other automaker can drive from on ramp to off ramp. They can even really change lanes.

1. Incorrect. Comparing human performance on _recognition_ over imagenet != comparing human total visual ability (object detection, which is a superset of recognition, + tracking) to a machine. Machines are far worse at accurately saying what + where things are + what they will do next. For example, cyclists move very quickly. Can your system not only see them but understand their intent with extremely high accuracy? What if your camera is occluded even partially, what failure mode do you have?

2. That's just wrong. If your image is washed out from low sun angle, all 200fps are washed out and you might misconstrue the red light for a green one.

3. You need n+1 for each kind of detector resilient to different failure modes to call yourself fault tolerant. 2 cameras both suceptible to the same problems isn't fully fault tolerant. Camera and radar have overlapping and joint uses but don't overlap enough e.g if radar fails camera can't do the same things the radar can.

4. This is unsubstantiated. You really think other sdc companies aren't collecting data from cars on the road? What do you think all those cars in mountain view and SF you see are for? What evidence do you have that Tesla is "looking at all cases" -- do they even have the infrastructure necessary to simulate their car's behavior over long-tail situations? (Hint: rumor is it's pretty lacking)

5. The serious L4+ companies e.g Uber are pretty good at driving highways. They're all working on the next phase of the problem.

5. See this is why I know you not actually reading the papers and seeing the comparison of their performance cause uber is doing terribly on the highway and otherwise. Raquel Urtasun already got her program shut down for 9 months for terrible performance. That has never happened to Tesla.

This is more due to regulatory games than performance. Tesla's aren't regulated like L4+ autonomous vehicles, so when they kill people (which they have!) they don't get shut down.

I would be scared to trust a self driving vehicle based on high resolution map: the government can change the roads any time, and I don't trust my government to notify Google in time about it.

,,You don’t ship mvp/beta of software that can kill people so you can gather training data.''

I think you haven't looked at the presentation: the new data gathering and verifying is based on shadow mode on model mispredictions, and it's not enabled until the misprediction rate goes down to an acceptable rate.

Reply on news.ycombinator.com