As has been detailed by @verygreen (and here: Tesla has allegedly activated a selfie cam Driver Monitoring System), there is a DMS being evaluated by Tesla now, using the in-cabin cameras on Model 3/Y (not clear whether this was already in the works or in response to European regulation and various recent "incidents"). It's not clear to me how much or what this is currently being used for (is it used to reduce nags while on AP? I don't know, and wasn't able to confirm reduction of nags or torque requirement in my vehicle - and I'm not sure I ever had the exact correct release, since I never saw it in the release notes for xx.15.xx - I'm really looking forward to the day when I can actually drive in AP with both hands firmly on the wheel again, haha.).
Do we think Tesla is going to soon prioritize safety by using this DMS output to dynamically adjust the well-known, hypothesized, "safety sliders?" They can do this specifically when the car is not in Autosteer/TACC mode, and on vehicles without FSD (but with HW3 of course). They currently do NOT mention this in the release notes.
Should be a reasonably easy thing to adjust the safety slider sensitivity upwards, when driver inattention is detected. And then adjust downwards again when the driver appears to be paying attention. Again, when NOT in Autopilot.
@verygreen recently posted a couple examples on Twitter which highlight this. It's a great point he is making (and one we have discussed here before, even without the DMS in place) - why not focus on improving safety by using the DMS to go into an ultra sensitive mode when inattention is detected? Why even worry about the nag threshold adjustment or adding extra nags, when in AP? The potential safety gains from the car intervening more aggressively when not using AP, when the driver is not paying attention, seem greater than those obtained from better monitoring while in AP... I would think these would improve Tesla's safety statistics quite considerably overall (though it would not show in their Autopilot statistics of course, and in fact it would probably close the gap to their selection-bias skewed AP statistics...though open the gap vs. all other cars). So why not? It's personally mysterious to me that Teslas with apparently fairly serviceable perception are still piling into vehicles by the side of the street. It seems unnecessary - even with false positives as a potential issue - especially with a DMS.
Or, do we think one or more of the following applies (the poll allows selection of multiple answers):
1) DMS not reliable enough yet to use as output for this slider adjustment (too many false positives for too sustained a period)?
2) Vehicle trajectory correction based on perception is inherently risky and Tesla doesn't trust the car to do it correctly (which might open them up to liability?), except in situations where AP is available and in use?
3) Vehicle perception is nowhere near reliable enough, too many false positives and false negatives, because of limitations of AI.
4) They're going to do it; they're just still working on it and it is going to come in the "immense foundational improvements" of V9. The difference will be gigantic.
5) The examples given above are with HW2.5 and the vehicle already does all of this and more with HW3, I'm just not trusting the car enough, and I should try running into parked cars to see what happens.
6) Tesla doesn't want to make their Autopilot statistics look relatively worse, which this would do.
7) The Tesla neural nets are learning from each of these collisions and it's just a matter of time before it starts avoiding collisions on its own.
8) This will lead to people driving around not paying attention, even if the car is beeping and braking and hassling them a lot, so Tesla doesn't want to encourage this.
9) Or?
This is the main reason I upgraded to HW3 (improved active safety features when not using AP/FSD), so hoping it happens ASAP!
Do we think Tesla is going to soon prioritize safety by using this DMS output to dynamically adjust the well-known, hypothesized, "safety sliders?" They can do this specifically when the car is not in Autosteer/TACC mode, and on vehicles without FSD (but with HW3 of course). They currently do NOT mention this in the release notes.
Should be a reasonably easy thing to adjust the safety slider sensitivity upwards, when driver inattention is detected. And then adjust downwards again when the driver appears to be paying attention. Again, when NOT in Autopilot.
@verygreen recently posted a couple examples on Twitter which highlight this. It's a great point he is making (and one we have discussed here before, even without the DMS in place) - why not focus on improving safety by using the DMS to go into an ultra sensitive mode when inattention is detected? Why even worry about the nag threshold adjustment or adding extra nags, when in AP? The potential safety gains from the car intervening more aggressively when not using AP, when the driver is not paying attention, seem greater than those obtained from better monitoring while in AP... I would think these would improve Tesla's safety statistics quite considerably overall (though it would not show in their Autopilot statistics of course, and in fact it would probably close the gap to their selection-bias skewed AP statistics...though open the gap vs. all other cars). So why not? It's personally mysterious to me that Teslas with apparently fairly serviceable perception are still piling into vehicles by the side of the street. It seems unnecessary - even with false positives as a potential issue - especially with a DMS.
Or, do we think one or more of the following applies (the poll allows selection of multiple answers):
1) DMS not reliable enough yet to use as output for this slider adjustment (too many false positives for too sustained a period)?
2) Vehicle trajectory correction based on perception is inherently risky and Tesla doesn't trust the car to do it correctly (which might open them up to liability?), except in situations where AP is available and in use?
3) Vehicle perception is nowhere near reliable enough, too many false positives and false negatives, because of limitations of AI.
4) They're going to do it; they're just still working on it and it is going to come in the "immense foundational improvements" of V9. The difference will be gigantic.
5) The examples given above are with HW2.5 and the vehicle already does all of this and more with HW3, I'm just not trusting the car enough, and I should try running into parked cars to see what happens.
6) Tesla doesn't want to make their Autopilot statistics look relatively worse, which this would do.
7) The Tesla neural nets are learning from each of these collisions and it's just a matter of time before it starts avoiding collisions on its own.
8) This will lead to people driving around not paying attention, even if the car is beeping and braking and hassling them a lot, so Tesla doesn't want to encourage this.
9) Or?
This is the main reason I upgraded to HW3 (improved active safety features when not using AP/FSD), so hoping it happens ASAP!