Quote:
Originally Posted by Seahawk
With your experience, can you recommend an overview of a system, or systems, to help folks like me catch up?
|
This overview is pretty good:
https://medium.com/data-science/a-crash-course-of-planning-for-perception-engineers-in-autonomous-driving-ede324d78717
My quick and dirty overview:
The autonomous driving "stack" is generally broken into a few modular areas which are sensor, perception, prediction, planning, and finally controls.
Sensors: hardware in the sensor array such as lidar, radar, camera, microphones, pressure sensors, GPS, etc etc
Perception: this is software which interprets the sensor data. Broadly speaking, this intake and interpretation of sensor data utilizes AI technologies that fall under the blanket of "
Computer Vision" which is the same branch of AI which is used to detect tumors and stuff like that. Perception answers questions like "what is that thing over there" and "is that traffic light green" and "is that a dog or a coyote."
Prediction: software which takes the perception as input and draws conclusions around future path of the objects (cars, pedestrians, street lights) identified by perception. Prediction answers questions like "where is that cyclist going to be in 1 second, 5 seconds, 20 seconds?" and "that car which I saw a second ago but which is now no longer visible because there is a bus between us, where do I think it is now?" and "that siren sound I am now hearing, what kind of emergency service is that and based on the layout of the buildings in this dense urban environment and the way the sound bounces off each surface which direction do I think it is coming from?"
Planning: software which takes the perception and prediction data as inputs and decides the best course of action for the autonomous vehicle. Planning answers the question "given the nearly infinite options available to me, what is the best one to set me up for success a millisecond from now, 10 seconds from now, 5 minutes from now?"
Controls: software and hardware, this is the actuation of the steering and the brakes etc. Baked into this is stuff like "my passenger and other drivers probably won't appreciate if I brake as hard as possible even if that is legal and logical thing to do otherwise."
Quote:
Originally Posted by masraum
If a human gets in an accident, that sucks, but doesn't seem crazy. If a machine gets in an accident, then that seems incomprehensible.
|
Yep, I feel this myself even if I know it is not logical... human drivers causing a crash every 500,000 miles is acceptable and autonomous drivers causing a crash every 9,000,000 miles is unacceptable.
Quote:
Originally Posted by LEAKYSEALS951
or guy who fell off his motorcycle suddenly bouncing at you(need to veer out of lane and into brakedown lane, etc...) How does the system evaluate and decide on a course of action on what seems to be an infinite number of scenarios?
|
This is a great use case for autonomous driving, because the vehicle is actively anticipating and planning all of these "what if" scenarios all the time. The vehicle sees the motorcycle rider and classifies them as a vulnerable road user, fundamentally different from a car or a dog or a Cessna 150. If the motorcycle rider becomes occluded (e.g. a larger vehicle comes between the autonomous car and the rider and for some time the rider is not visible) the autonomous vehicle continues to track/remember that a motorcycle is over there, and actively anticipates the likelyhood of where and when they will next be visible and what their vectors and actions will be. If the rider suddenly comes off their bike the autonomous vehicle doesn't need to think about what to do... it already has "what if" plan and it puts that plan into action instantly and is cool as a cucumber the whole time.
Overview of the different "levels" of autonomous vehicles as defined by SAE (Tesla is level 2, Waymo/Zoox/Stack AV/Aurora are level 4)
https://en.wikipedia.org/wiki/Self-driving_car