7 ms·

As an ML engineer I'm a little bit baffled that Tesla has not solved this by now. It's not like they lack data or ML knowledge. It seems like they should have

• AceJohnny2 5y ago
  Years ago, a friend worked in the Autopilot group. It took them a year to procure servers to store the telemetry of the existing cars, and then weeks to have them setup. They don't work there anymore. From their experience, I know one thing: I will never work for Elon Musk. He may be a great visionary and salesman, but he's a horrible manager.
  • etaioinshrdlu 5y ago
    This is fascinating, but what caused the delays specifically?
    • AceJohnny2 5y ago
      Server procurement was a CFO thing, and their specific requirements didn't fit under the existing buckets, so it took them a long time to get it approved. It was stunning to see the complete disconnect between Musk's grand declarations and what the organization was actually setup to deliver. Frankly, it just gives me more respect for Tim Cook, who as COO at Apple made his company able to turnaround and deliver HW in record time. Edit: in retrospect I wonder if Musk's grand public declarations were actually a way to control and pressure his own organization. Remember, Musk didn't actually found Tesla, he rescued it from bankruptcy after the Roadster didn't return as much as needed, so he inherited an existing structure.
  • systemvoltage 5y ago
    Just like your anecdote, I have one to share as well: I know a close friend that works on space lasers at SpaceX and has a blast, best job ever according to him and he thinks Elon is an excellent manager mostly because there is zero bureaucracy and people are not afraid of "Do nothing" option as well as removing complexity. In fact, he wouldn't work anywhere else after seeing the company culture. I know it's cool to hate Elon.
    • ipsum2 5y ago
      I don't think Elon has a day-to-day role at SpaceX unlike at Tesla, right? From what I've read, Gwynne Shotwell is the main person in charge there.
      • simondotau 5y ago
        Elon spends nearly as much time at SpaceX as he does Tesla. There's no question that Elon regularly mucks in at the lowest levels of engineering at SpaceX. If you want evidence from a reasonably neutral observer, take Sandy Munro (himself an engineer who has worked on everything from cars to aeroplanes). He recently interviewed Elon, ostensibly about Tesla but the interview was in a meeting room at SpaceX. After the interview he was invited to a two hour design review meeting and was "blown away" at Elon's depth of involvement. https://youtu.be/S1nc_chrNQk?t=370 https://youtu.be/S1nc_chrNQk?t=370 (6:10 to 8:45)
      • Geee 5y ago
        He is very much involved at actual day-to-day engineering at SpaceX. He is the CEO and the chief engineer there. Sources: https://www.reddit.com/r/SpaceXLounge/comments/k1e0ta/evidence_that_musk_is_the_chief_engineer_of_spacex/ https://www.reddit.com/r/SpaceXLounge/comments/k1e0ta/eviden... https://www.youtube.com/watch?v=S1nc_chrNQk&t=377s https://www.youtube.com/watch?v=S1nc_chrNQk&t=377s
    • AceJohnny2 5y ago
      I dunno, "move fast and break things" works for rockets [1], not so much for cars. [1] https://youtu.be/bvim4rsNHkQ https://youtu.be/bvim4rsNHkQ
• Syonyk 5y ago
  > I think the real problem is obvious. Yes. The human brain and visual systems aren't nearly so trivial to replicate as a lot of people in the tech industry seem to think. Tesla is just one of many case studies in the paired tech industry arrogance seen so frequently: - "A human is just a couple really crappy cameras and a neural network, we know how to do better cameras and neural networks, how hard can it be?" - "We can do anything we dream with 99.995% reliability in the synthetic, computer-based world of the internet because we know code. Therefore, we can do anything we want in the physical reality with code!" Both are far from evident in practice, but the belief in them continues, despite it being increasingly obvious to everyone else that neither one is true. Human vision and world processing is quite impressive - and, as pointed out elsewhere in this thread, a two or three year old would have no trouble working out that the obstacles were some things on a truck. I've got a nearly three year old, and I guarantee he wouldn't confuse those for stoplights in the slightest. I also wouldn't let him out on the road, though he does well enough with a little Power Wheels type toy. But there is far more going on in the visual processing system than we even understand yet, much less have the slightest clue how to replicate. And while code may be fine on the internet (where you can retry failed API calls and things mostly make sense), the quote about how fiction is constrained by what's believable and reality sees no such restrictions is very true. Out on the roads, all sorts of absolutely insane things can and do happen on a regular basis - and you can't predict or plan for all of them. But the car has to handle them or it crashes. As a random example, a year or two ago, I was behind a car that had poorly strapped a chunk of plywood to their roofrack with a good chunk hanging forward, and the front end of it was starting to oscillate awfully hard. I had a good clue that it was going to come apart sometime in the very near future, so backed off from a normal following distance to quite a way back. Sure enough, half a mile later, it failed, went flying through the air, slammed into the road a good distance behind the car, and tumbled a bit. Had I been using a normal in town following distance, it would have either hit me or tumbled into me, but using a human visual system, it was obvious that my existing following distance stood a good chance of being a bad idea. Stuff like this happens on roads constantly. Meanwhile, state of the art self driving can't tell the difference between stoplights and some poles on a truck. You'll excuse me if I don't think the problem is anywhere remotely close to solved for a general case Level 4 purpose.
  • 5y ago
• fsh 5y ago
  I've never understood this argument. Isn't the main bottleneck that you need well-labeled data for training the neural networks? How is having tons and tons of random camera footage going to help?
  • ampdepolymerase 5y ago
    For a lot of problems, clean data labeled data is no longer the bottle neck, or in some cases, there are ways around it. The bigger issue is dealing with the "long-tail" of unknown new scenarios. This is a currently unsolved challenge.
  • unoti 5y ago
    > How is having tons and tons of random camera footage going to help? One way to think of this is that that the footage is implicitly labelled: we have the benefit of hindsight: we know what the state/location of the vehicle was going into the future. That benefit of hindsight also can serve as implicit labels by knowledge that the vehicle did not crash or collide with something immediately after the footage.
  • stonemetal12 5y ago
    Fairly simple, they have footage from every time a person had to correct AP's driving. Take that footage label it. Train some of it, save some for test cases. Finally don't release an update until it drives better than the current system.
• ModernMech 5y ago
  Right? This should be trivially solved with some sort of temporal filtering on detected objects. If you detect a traffic light at (x,y) in one frame, and it disappears in the next, but there's a new one at (x+dx,y+dy), then you shouldn't place a new one down in the world frame. You should only place a traffic light down if you're confident it exists and is operational. At the very least, the lights should be detected on the back of the truck, but they should move with the truck. At least that matches what's happening. I don't understand why this is hard for Tesla engineers -- I was doing this kind of thing in grad school a decade+ ago and it worked fine. I've seen it in other demos where object classifications rapidly cycle between person, bike, car, etc. Are they not filtering anything? Is this a symptom of "AI-ing all the things"? Because we did it with bog standard computer vision techniques back then and never got behaviors like this.
  • nmca 5y ago
    Did any of your grad school work make it into the real world at all? Typically I'd suggest that such ideas are simple in theory and difficult in practice.
    • ModernMech 5y ago
      Yes, there are real systems out there working off of the techniques we used back then. I didn't work much in theory at all.
• kempbellt 5y ago
  I'm curious what the issue is here that seems unsolved? It's unconventionally displaying what it is recognizing, but the car isn't doing anything janky. The car is properly recognizing traffic lights pretty darned well, considering the circumstance. It looks like it has a built in understanding that traffic lights are "always" stationary - hence, assigning them static locations on the 3D map - but it keeps having to update the model because the lights are actually moving. This seems like a very non-obvious edge case that I wouldn't expect an ML team to even consider as a possibility. Now they need to program into the ML model an understanding that traffic lights are typically stationary. Which seems even more difficult to me, from a technical perspective - you don't want false negatives... The car isn't braking or making any strange maneuvers from what I can tell. I'm actually impressed that it's handling it this well.
  • sabhiram 5y ago
    So you are impressed that they ran headfirst into a bunch of traffic lights because they did not know what to do? When you don't know what to do - do nothing. What if it was a traffic light on roller skates? Or a kid, dressed as a traffic light, on roller skates?
    • codeulike 5y ago
      The car is not under autopilot. The driver is driving. (the grey steering wheel icon would be blue if autopilot was on)
    • comradesmith 5y ago
      Stopping would be a form of action too. ¯\_ツ_/¯
    • kempbellt 5y ago
      I don't know exactly how Tesla's safety systems are designed, but this is my guess. Collision detection systems (radar) are accurately not detecting an impending collision because the lights are not actually on a collision course with the vehicle. Object recognition systems (computer vision) are working very well, because they recognize the lights and are updating the 3D map accordingly, but the traffic light 3D model is not designed to be a moving object - unlike vehicles, which frequently move. Which is why we see the car "passing through" them. What we are likely seeing is simply a weird edge case in the output for the user-interface. I'd imagine if an object was actually flying at the car and the car could see it, it would brake accordingly. Also, the map is two-dimensional. The car frequently drives underneath traffic lights that I'm sure also appear "on top of" the car in normal cases. Object recognition and collision detection, from what I understand, are two very different systems.
• Havoc 5y ago
  > they should have a million hard test cases that must pass Move fast and break things [like tests]
• visarga 5y ago
  As another ML engineer I have all the respect for Andrej Karpathy and wouldn't dare to walk in his shoes. My ML model could kill people? No thanks, I couldn't sleep anymore.
  • des1nderlase 5y ago
    That's reason more to use synthetics to generate and train on many many edge cases. Getting photo realistic scenes is easier than ever.