Hacker News Reader

Comment by whatever1 | Hacker News Reader

And not even one accelerometer. Will definitely work safely /s.

I'm almost certain that nearly everyone uses accelerometers, gyroscopes and wheel odometry data from the base vehicle. These sensors are dirt cheap and kinda hidden inside the vehicles so people don't write articles about them. That doesn't mean that engineers forgot about them, they are just not what makes a journalist's heart pump faster.

The article doesn't use a consistent definition of sensors. Notice how Mercedes' number is reported to include their emergency microphone as a sensor, but Tesla's number only includes cameras + ultrasonics. Ultrasonics are counted, but not the IMUs that may also be involved in detecting collisions.

If you were to truly count every sensor connected to every chip, there'd be hundreds or thousands of sensors. Only some of them are used. Different vehicles in a testing fleet will have slightly different numbers as well, due to the various component changes that are in-flight at any given time.

What kind of useful data would accelerometer provide? 99.999% of the time wheel rotation rate matches exactly car movement speed and direction, so as long as you have odometer (resolver/encoder), you have accelerometer data for free. For the cases when wheels are spinning freely, you have a serious situation which needs to be dealt with using specialized safety equipment.

The accelerometer exactly matches wheel rotation rate when you are driving in a straight line on a planar surface[1] in a vehicle with no toe. This is rarely the case: roads aren't perfectly flat, and outside of North America, aren't even remotely straight in most circumstances, and the handling characteristics of 0 toe are terrible. Granted, the inaccuracy from non-planarity of the road is minimal, but building an accurate enough model of tire slip from steering geometry is hard, and changes based on the level of wear on the tires, their temperature, and the road conditions, and the deviation between acceleration and wheel telemetry adds up very quickly.

So yes, you absolutely do need an IMU for self-driving.

Fun thing though, VW implemented tyre pressure monitoring in their vehicles with just a firmware update by simply comparing wheel rotation speeds in the ABS ECU. If one wheel is consistently faster than the rest, it's got a smaller radius, most likely because the tyre is flat.

First of all, pose. Secondly, different transfer functions. A wheel is a heavy physical thing that takes time to change. An accelerometer isn't, by design. The latter also redundantly measures acceleration direction rather than requiring it as an input. Thirdly, encoders don't always match accelerometers even in normal driving. Turns at intersections with tram tracks commonly cause these sorts of errors.

Haha. Have you ever calculated movement from wheel speed sensors? On top the other sensors are mounted on the body and there is a whole world of mass-spring and tire equations between... And then there is parking with the nitty gritty details of tire models etc...

Funny. Just today: https://imgur.io/gallery/hMpaVzC

The nice thing about sensor fusion is that throwing more information into it generally always makes things better (so long as you have a correct idea of the reliability of it). IMUs are absolutely going to be used to feed data into the localisation of basically every system, but it's so basic it doesn't really bear mentioning in the context of the other sensors you're using.

Accelerometers would tell you if the wheel rotation and steering angle is still matching the movement of the vehicle.

This is useful for detecting stability problems like over- or understeer.

If the steering is pointing slightly left, the car is moving rapidly right, and both rear wheels are turning at roughly the same speed and way faster than the fronts, you've lost traction and are oversteering like crazy.

Reply on news.ycombinator.com