3 ms·

Part of the fix was to use radar: https://www.tesla.com/blog/upgrading-autopilot-seeing-world-radar https://www.tesla.com/blog/upgrading-autopilot-seeing-world-

Part of the fix was to use radar: https://www.tesla.com/blog/upgrading-autopilot-seeing-world-radar https://www.tesla.com/blog/upgrading-autopilot-seeing-world-...

studentrob 9y ago
Oh interesting. That's a pretty detailed report. One thing jumps out at me, > This is where fleet learning comes in handy. Initially, the vehicle fleet will take no action except to note the position of road signs, bridges and other stationary objects, mapping the world according to radar. The car computer will then silently compare when it would have braked to the driver action and upload that to the Tesla database. If several cars drive safely past a given radar object, whether Autopilot is turned on or off, then that object is added to the geocoded whitelist. Relying on whitelists seems like a hack. Then again, I'm not building it =)
- ramdomize 9y ago
  > Relying on whitelists seems like a hack. Then again, I'm not building it =) Agree that it feels hacky at first, but once the whitelist dataset gets huge, it becomes unique training data for Tesla (and hopefully their machine learning will be able to generalize from it).
  - studentrob 9y ago
    > Agree that it feels hacky at first, but once the whitelist dataset gets huge, it becomes unique training data for Tesla (and hopefully their machine learning will be able to generalize from it). I guess. But then you still have to deal with rollout in other countries. And maintenance of sign locations which can change over time, get removed, etc. Still hacky IMHO. How would machine learning make use of white-listed data? I doubt they could use that data to predict the GPS location of unknown signs. If you mean image recognition, I assume if machine learning could properly identify the signs with accuracy, then they wouldn't need the whitelist. Then again, maybe they truly haven't collected a full overhead-sign dataset yet. I'd be shocked though if they don't by now. Anyway, you could be right. It would be fun to learn more about these setups.

Part of the fix was to use radar: https://www.tesla.com/blog/upgrading-autopilot-seeing-world-radar https://www.tesla.com/blog/upgrading-autopilot-seeing-world-...

studentrob 9y ago

Oh interesting. That's a pretty detailed report. One thing jumps out at me, > This is where fleet learning comes in handy. Initially, the vehicle fleet will take no action except to note the position of road signs, bridges and other stationary objects, mapping the world according to radar. The car computer will then silently compare when it would have braked to the driver action and upload that to the Tesla database. If several cars drive safely past a given radar object, whether Autopilot is turned on or off, then that object is added to the geocoded whitelist. Relying on whitelists seems like a hack. Then again, I'm not building it =)

ramdomize 9y ago
> Relying on whitelists seems like a hack. Then again, I'm not building it =) Agree that it feels hacky at first, but once the whitelist dataset gets huge, it becomes unique training data for Tesla (and hopefully their machine learning will be able to generalize from it).
- studentrob 9y ago
  > Agree that it feels hacky at first, but once the whitelist dataset gets huge, it becomes unique training data for Tesla (and hopefully their machine learning will be able to generalize from it). I guess. But then you still have to deal with rollout in other countries. And maintenance of sign locations which can change over time, get removed, etc. Still hacky IMHO. How would machine learning make use of white-listed data? I doubt they could use that data to predict the GPS location of unknown signs. If you mean image recognition, I assume if machine learning could properly identify the signs with accuracy, then they wouldn't need the whitelist. Then again, maybe they truly haven't collected a full overhead-sign dataset yet. I'd be shocked though if they don't by now. Anyway, you could be right. It would be fun to learn more about these setups.