... working on this stuff day and night ... —- in California.
I’ve been driving for 40+ years, millions of km, 1/2 during winter.
Now, Elon is a Saskatchewan boy, so he should know better. Perhaps he’s forgotten. Winter driving is at best 1/2 vision dependent, often much less. Many times, you have to
actively ignore your vision (everything you see is “moving” sideways). Quite often, you’re modulating your throttle to maintain a tiny ratio of +/- acceleration that maintains static friction; as soon as 1 or more contact patches achieve dynamic friction, you’re entering a spin due to yaw forces induced by your other driven wheels.
Much of the time, the road surface adhesion characteristics are best detected by sheen, vibration, sound and guesswork/prediction based on temperature, sunlight, recent weather, etc.
Unless the vehicle has sound, vibration, and vision sensor integration — it cannot hold a candle to human driving, except in the most trivial driving scenarios.
Winter is a complete non-starter for any autonomous driving technology that I am aware of.
I don’t know if Lidar would really help much. It can’t punch through heavy rain/snow, at least not as well as radar anyway. Maybe it would help in trivial driving conditions. I think vision, sound, vibration, and radar with a powerful ANN trained by professional all-season winter drivers in challenging conditions might be useful, and eventually even good. It should be able to perform super-challenging evasion and recovery maneuvers that could save lives!
The vision problems are real. But the ability to handle ice/snow-induced wheel spin and slew seems like exactly the kind of problem that can be handled with sensitive electronic traction control, perhaps augmented with accelerometers. That specific problem doesn't seem intrinsically more complicated than, say, keeping a drone level on a breezy day -- which is something machines already do much more reliably than humans can. (New England driver here, so I do have some personal experience.)
Traction/stability control can only help with the situation you are already in. Being able to see the condition of the road -- the sparkle of ice forming, for example, or a sheen that suggests black ice, etc -- allows you to act instead of just react.
While vision is a hard problem in general, the specific challenge of "ice detection" seems like exactly the kind of problem that computers can be programmed to do better (and more tirelessly) than humans, particularly with specialized detectors and LIDAR. Better sensors, better reaction time, zero distractions. (I don't have any expertise here, just intuition.)
With the right instruments, I don't disagree. But in the context of cameras, I think it will be quite some time before that happens. Watching demos of state-of-the-art computer vision object detection today, I think it may be decades before we have cameras and computers good enough to do the kind of on-the-fly visual analysis that humans do naturally.
> in the context of cameras, I think it will be quite some time before that happens.
Why? The car can easily detect slipping. It can upload camera data every time any Tesla ever lost traction. You think machine learning won't be able to correlate those signals?
Certainly some problems are way easier for us, but that problem seems way easier for the Tesla fleet than for a human.
When the first snowstorm hits it never ceases to surprise me how the dynamics of the road change - not just the friction, but the number of lanes and where they go all change as people give their best guess, which over the next 2000 cards becomes codified. If you have self-driving cars following the old lanes you're going to have accidents.
Dunno, I think of the people that follow other cars as sheep. In my subaru I can pick any lane I like, and often the one with the least travel has the best traction. With snow tires and a subaru I can handle just about any conditions up to where the snow lifts up the car off the tires.... which I've had happen. Digging compressed snow out from under a car is no fun.
So often I see a giant line of cars behind the plow, slipping and sliding along, then I pick the deepest snow lane with the most fresh snow and it's just fine.
Generally it's not the lane you pick, but the safe speed for that lane. Sure that might vary per lane, but with today's sensors and electric motors the autonomous driver should be more accurate at quantifying that than a human. After all a computer could easily say apply 50HP to each wheel (one at a time) for 10ms or similar to quantify where the dynamic/static threshold is.
It has nothing to do with the traction, it's the fact that people are now travelling across the road at different angles than they were before, if you're trying to stay in the regular lane, you're going to be fighting people moving across the lane.
This is silly. The car has a steering wheel. It won’t engage autopilot in a situation that it can’t handle. Just like waymo cars only drive in Phoenix.
I’ve been driving for 40+ years, millions of km, 1/2 during winter.
Now, Elon is a Saskatchewan boy, so he should know better. Perhaps he’s forgotten. Winter driving is at best 1/2 vision dependent, often much less. Many times, you have to actively ignore your vision (everything you see is “moving” sideways). Quite often, you’re modulating your throttle to maintain a tiny ratio of +/- acceleration that maintains static friction; as soon as 1 or more contact patches achieve dynamic friction, you’re entering a spin due to yaw forces induced by your other driven wheels.
Much of the time, the road surface adhesion characteristics are best detected by sheen, vibration, sound and guesswork/prediction based on temperature, sunlight, recent weather, etc.
Unless the vehicle has sound, vibration, and vision sensor integration — it cannot hold a candle to human driving, except in the most trivial driving scenarios.
Winter is a complete non-starter for any autonomous driving technology that I am aware of.
I don’t know if Lidar would really help much. It can’t punch through heavy rain/snow, at least not as well as radar anyway. Maybe it would help in trivial driving conditions. I think vision, sound, vibration, and radar with a powerful ANN trained by professional all-season winter drivers in challenging conditions might be useful, and eventually even good. It should be able to perform super-challenging evasion and recovery maneuvers that could save lives!