3 ms·

In the article I posted it mentions that of the 56% of pilots who admitted falling asleep, 29% also admitted to having woken up to a sleeping copilot. There is

In the article I posted it mentions that of the 56% of pilots who admitted falling asleep, 29% also admitted to having woken up to a sleeping copilot. There is no path from Autopilot as it is to full autonomy without doing a major shift in strategy. Google figured out very early on that the incremental approach is unfeasible. Most of the other big car companies at the outset were intending to develop autonomy incrementally as well, but then they thought about it for a few minutes.

Robotbeat 9y ago
Why aren't major shifts in strategy possible? LIDAR may become very cheap (i.e. $1000, much less than drivers paid for the full self-driving package). Improvements in radar may also help. Requiring much lower speeds (driving like a Waymo grandma) and specific conditions (i.e. dry, well-lit, well-mapped) are also feasible. As the technology improves, conditions can be relaxed for full self-driving. The very fact that in this case, the driver had multiple problems with this same spot, and that other drivers also noticed problems in this area, points to an obvious mitigation strategy: flag problem areas geographically for autopilot, develop specific strategies that address the problem, and verify it has been fixed by examining how other Teslas on the road in "ghost" mode (i.e. with the human driving but the software pretending to self-drive) respond to the changes. Tesla, due to its massive and well-connected fleet, has a lot of tools in the toolbox to address these problems, in some cases more than Google (although I agree Waymo's self-driving capability is currently more impressive).
- sah2ed 9y ago
  > flag problem areas geographically for autopilot, develop specific strategies that address the problem, and verify it has been fixed by examining how other Teslas on the road in "ghost" mode This is what I have been thinking as well, except it appears that Tesla doesn't currently have a way to handle such scenarios. Ideally, they should be recording when a driver makes a sharp swerve to correct autopilot's naïve lane marker following algorithm, then use that information to inform future drives using autopilot on that stretch by disabling autopilot as that patch of road is approached, while they work on a reliable fix.
  - Robotbeat 9y ago
    Tesla's new Waze-like real-time traffic system as part of its mapping software update could help handle this. If the autopilot system (engaged or in ghost mode) generates an error signal (either binary or some variable degree of error) that is pushed to the traffic system along with the traffic data, it could easily alert other drivers (and hopefully the autopilot system) of such problems. Considering they have enough data to do realtime traffic analysis, they easily should have enough data to determine problem areas for the autopilot.

Robotbeat 9y ago

Why aren't major shifts in strategy possible? LIDAR may become very cheap (i.e. $1000, much less than drivers paid for the full self-driving package). Improvements in radar may also help. Requiring much lower speeds (driving like a Waymo grandma) and specific conditions (i.e. dry, well-lit, well-mapped) are also feasible. As the technology improves, conditions can be relaxed for full self-driving. The very fact that in this case, the driver had multiple problems with this same spot, and that other drivers also noticed problems in this area, points to an obvious mitigation strategy: flag problem areas geographically for autopilot, develop specific strategies that address the problem, and verify it has been fixed by examining how other Teslas on the road in "ghost" mode (i.e. with the human driving but the software pretending to self-drive) respond to the changes. Tesla, due to its massive and well-connected fleet, has a lot of tools in the toolbox to address these problems, in some cases more than Google (although I agree Waymo's self-driving capability is currently more impressive).

sah2ed 9y ago
> flag problem areas geographically for autopilot, develop specific strategies that address the problem, and verify it has been fixed by examining how other Teslas on the road in "ghost" mode This is what I have been thinking as well, except it appears that Tesla doesn't currently have a way to handle such scenarios. Ideally, they should be recording when a driver makes a sharp swerve to correct autopilot's naïve lane marker following algorithm, then use that information to inform future drives using autopilot on that stretch by disabling autopilot as that patch of road is approached, while they work on a reliable fix.
- Robotbeat 9y ago
  Tesla's new Waze-like real-time traffic system as part of its mapping software update could help handle this. If the autopilot system (engaged or in ghost mode) generates an error signal (either binary or some variable degree of error) that is pushed to the traffic system along with the traffic data, it could easily alert other drivers (and hopefully the autopilot system) of such problems. Considering they have enough data to do realtime traffic analysis, they easily should have enough data to determine problem areas for the autopilot.