4 ms·
Andrej Karpathy already explained they were getting too much noise from Radar. Adding LIDAR would only complicated it. Unless you have better knowledge than A
by TheParkShark 5y ago
Andrej Karpathy already explained they were getting too much noise from Radar. Adding LIDAR would only complicated it. Unless you have better knowledge than Andrej Karpathy, I’ll go with his word on this.
Edit: nitrogen I can’t comment directly on yours, so I’ll ass it here. That’s a good goal to have and they tried that for well over two years. They found too much noise and confusion. If Radar says one thing and Vision says another over and over who do you trust? They decided they could train their NN with just Vision. They use LIDAR on test cars to confirm their Vision can recognize distances correctly. In this case I believe cutting out a sensor instead of trying to neuter it is the right decision.
- nitrogen 5y agoIsn't the basic goal of sensory fusion to be able to use each sensor to the extent that it's good, and ignore it to the extent that it's bad?
- qeternity 5y agoKarpathy is the guy responsible for the results we see in these videos. Why would you trust his word about any of this?
- sbierwagen 5y agoAndrej works for Tesla, so yes, I don't believe him. If he contradicted his boss in public, Musk would just fire him and hire someone who will repeat the party line. I won't believe anything he says about lidar until he's no longer at Tesla, or until they have a car that's not trying to ram concrete pillars.
- londons_explore 5y agoAndrej has said some stuff in public that certainly implies he's prioritizing the 'party line' on stuff like HD maps and lidar than his academic stats knowledge.
- nuker 5y ago> Adding LIDAR would only complicated it. More sensors is better. They just have not yet learned how to merge data. Example is drone autopilot, that have IMU and GPS and uses Kalman filtering.
- tw04 5y ago> If Radar says one thing and Vision says another over and over who do you trust? Neither? If you don't have a tiebreaker you disengage and let the human decide. We solved this problem in systems long ago. You either fail open or fail closed depending on the system. In this case it seems like you'd always fail to the safest option which is alert the human.
- soco 5y agoSo basically you're driving full speed against an obstacle and you get the commands handed out. It sounds like a slapstick comedy where the pilot says "lol bye you're on your own" and jumps out the plane door.
- kfarr 5y agoI think that is still desirable behavior when compared to a slapstick comedy where the pilot says "don't worry bruh I got this" and then proceeds to plow directly into a mountain side
- soco 5y agoI get your point and I agree to it. For the nitpicking, I'd just not use the word "desirable" when the slapstick situation would be rather "less bad"... because "desirable" means a positive outcome while neither one is anything positive.
- drcongo 5y agoI learned on HN long ago that planes usually have three sensors for critical systems so that if one is disagreeing but the other two agree, you still have a reasonable degree of certainty. Having a single sensor that seems to insist on steering toward obstacles doesn't seem great to me.