So the biggest news here seems to be the upgrades to Autopilot. (https://www.tesla.com/blog/upgrading-autopilot-seeing-world-...). They describe a fantastic amount of new technology, including using radar to make a point cloud despite the fact that "[s]omething made of wood or painted plastic ... is almost as transparent as glass to radar." This venture reeks Murphy's Law. I'll let someone else be the guinea pig in this experiment.
"It is hard to tell from a single frame whether an object is moving or stationary or to distinguish spurious reflections."
What, they're not getting range rate directly from the radar? You can tell if an object is moving towards or away from you from one frame if you have range rate. Bosch automotive radars do return range rate.[1] Even 1990s automotive radars returned range rate. That's one of the huge advantages of radar over vision. Range and range rate go into a very simple formula which yields time to collision. Vision is lousy at range rate and gets worse as the distance increases. Radar is good at range rate and the error is constant out the range of the radar.
Whomever wrote that piece of PR probably doesn't really know how it works.
My recollection is Tesla has been leaning heavily on optical image processing. Perhaps the radar is, or was, not very sophisticated? For example, can a single unit by itself tell, in one frame, if an object is moving laterally to its field of view?
The radar was added to all Tesla vehicles in October 2014 as part of the Autopilot hardware suite, but was only meant to be a supplementary sensor to the primary camera and image processing system.
New autopilot updates are only pushed out after they've been run against simulated data (in house validation, "the gauntlet" corpus of scenarios) and then real world data (inertly) and perform significantly better than existing algorithms.
Presumably there are a ton of beta testers that put a 100,000 or so miles running this actively in a wide array of road driving scenarios before they roll this out to customers as well. Under beta testing scenarios, a single driver should be, even with time for charging, be able to do about 250 miles/day of driving - figure 10,000 miles/month. So, to get 100K of real-world road experience, would only take 10 drivers/month (or 20 drivers 2 weeks, etc.... Mythical Man month doesn't apply here)
Tesla is accumulating ~1 million miles of data every 10 hours fleet-wide [1]. With their environment, they could presumably beta test across the entire fleet, as the changes are inert and only data is sent back as to what the vehicle would've done (compared to what it actually did).
Understood they do inert testing with their existing fleet. I'm talking about real, rubber hits the road, vehicles actively running the firmware. I have to believe that prior to rolling out firmware on the fleet, Tesla rolls it out to be actively run on some set of beta testing vehicles - perhaps actively run by Tesla Employees driving a variety of real-world road conditions.
Correct, but you need to sign an NDA to run (or rather, for your vehicle's VIN to be whitelisted for pre-release firmware bundle retrievals) a non-GA release.
Actually, I think they record all of the data all of the time anyway, so they can turn the sensor on and turn off the decisionmaking, and gather info using real-world data to see what the new version would have done.
It's like beta-testing without actually beta-testing.
Comments
So the biggest news here seems to be the upgrades to Autopilot. (https://www.tesla.com/blog/upgrading-autopilot-seeing-world-...). They describe a fantastic amount of new technology, including using radar to make a point cloud despite the fact that "[s]omething made of wood or painted plastic ... is almost as transparent as glass to radar." This venture reeks Murphy's Law. I'll let someone else be the guinea pig in this experiment.
"It is hard to tell from a single frame whether an object is moving or stationary or to distinguish spurious reflections."
What, they're not getting range rate directly from the radar? You can tell if an object is moving towards or away from you from one frame if you have range rate. Bosch automotive radars do return range rate.[1] Even 1990s automotive radars returned range rate. That's one of the huge advantages of radar over vision. Range and range rate go into a very simple formula which yields time to collision. Vision is lousy at range rate and gets worse as the distance increases. Radar is good at range rate and the error is constant out the range of the radar.
Whomever wrote that piece of PR probably doesn't really know how it works.
[1] https://etd.ohiolink.edu/rws_etd/document/get/ohiou130408338...
My recollection is Tesla has been leaning heavily on optical image processing. Perhaps the radar is, or was, not very sophisticated? For example, can a single unit by itself tell, in one frame, if an object is moving laterally to its field of view?
The radar was added to all Tesla vehicles in October 2014 as part of the Autopilot hardware suite, but was only meant to be a supplementary sensor to the primary camera and image processing system.
"I'll let someone else be the guinea pig in this experiment."
You don't actually have this option - if you are sharing roads with these cars, you're automatically a beta tester.
Nice thought, but they're on the road with everyone else, so it's kind of not our choice about who gets to be subject to this experiment.
New autopilot updates are only pushed out after they've been run against simulated data (in house validation, "the gauntlet" corpus of scenarios) and then real world data (inertly) and perform significantly better than existing algorithms.
https://youtu.be/2vHO1p9717s?t=8m45s
Presumably there are a ton of beta testers that put a 100,000 or so miles running this actively in a wide array of road driving scenarios before they roll this out to customers as well. Under beta testing scenarios, a single driver should be, even with time for charging, be able to do about 250 miles/day of driving - figure 10,000 miles/month. So, to get 100K of real-world road experience, would only take 10 drivers/month (or 20 drivers 2 weeks, etc.... Mythical Man month doesn't apply here)
Tesla is accumulating ~1 million miles of data every 10 hours fleet-wide [1]. With their environment, they could presumably beta test across the entire fleet, as the changes are inert and only data is sent back as to what the vehicle would've done (compared to what it actually did).
[1] http://qz.com/694520/tesla-has-780-million-miles-of-driving-...
Understood they do inert testing with their existing fleet. I'm talking about real, rubber hits the road, vehicles actively running the firmware. I have to believe that prior to rolling out firmware on the fleet, Tesla rolls it out to be actively run on some set of beta testing vehicles - perhaps actively run by Tesla Employees driving a variety of real-world road conditions.
Correct, but you need to sign an NDA to run (or rather, for your vehicle's VIN to be whitelisted for pre-release firmware bundle retrievals) a non-GA release.
Actually, I think they record all of the data all of the time anyway, so they can turn the sensor on and turn off the decisionmaking, and gather info using real-world data to see what the new version would have done.
It's like beta-testing without actually beta-testing.