As far as AI is concerned, a trolley on a track a way simpler problem than driving a car
- no need to detect or infer the edge of the roadway
- all road will be familiar and pre-programmed down to the mm
- all track and road signals will be exactly known (it will either know electronically, or will know where to look and when a look was insufficient)
- GPS + track location = super accurate awareness
- no computation devoted to staying in lane or planning evasive maneuvers
- no speeds over 50 (ish?) (limits how far ahead to consider)
Between "watching videos" and several years (across all seasons and conditions) of co-operation with humans, the system will know every little bump and feel of every route, even wrong-railing, and base its actions as a rookie on what amounts to hundreds of years of operations.
The vehicle will know the exact place it will be "now" and "seconds from now"...the only question will be how quickly (and at what speed) it will arrive "there" and has it detected and tracked all people/objects that might be "there"
It is the "forward collision avoidance" problem (which Subarus have been really good at since 2012) both straight ahead, and tracking objects in motion into that path from any other direction. Add a few more sensors, and the train will be "aware" of all interactions near the train within inches in every direction.
The system will likely know that that BU student is drunk just by looking at their gait, and will see that a crossing car is going too fast to stop (by knowing its exact speed and direction)
(As noted, Uber had told its killer car to treat unexpected objects as if they were a cardboard box or plastic bag blown in--this was a programming failure not a technical barrier)
And it will not get tired or distracted.
Perhaps, as on the Docklands Light Railway, a human will be on board to answer questions and provide a sense of security (or to check fares), but the main value a human will add will not be the ordinary operation of a tracked vehicle and the anticipation of nearby events/interactions/objects.
Economically, the highest paid drivers are at greatest risk first because their high wage rate makes it easier to justify expensive sensor packages (LIDARs and things to infer density (life forms) vs empty shapes (boxes and bags).
Uber drivers will come next because eventually computers will underprice them as "employees" Your personal car will come last (except it already has FCW and AEB, which are most of the game for tracked vehicles) because it still has to be priced as a car, not an economic tool.
AEB = Automatic Emergency Braking; Just like it sounds, the car will apply the brakes without driver input when an accident is likely. Relies on a Forward Collision detection, warning, or avoidance sensor.
FCW = Forward Collision Warning; the ability to do the math (continuously) based on the position and velocity of objects to project collisions in upcoming seconds. When a collision is of high enough likelihood, it can warn to mitigate or avoid the accident by braking.
FCW has progressed rapidly. From silent, to audible, to automatic. Taking the typical Toyota as an example:
2010 = Silent. Systems would not actually warn the driver but would warn other systems: boosting braking power, pre-tension seat belts, hoping the driver would act. Some would brake only after an accident was unavoidable.
2015 = systems would sound an alert in time for a human to stop the accident, but if ignored, only brake to mitigate impact, not avoid. Goal: reduce the car's speed at impact.
2018 = systems would both warn and trigger AEB in time to *avoid* an accident in most cases
LIDAR is *not* needed for FCW and AEB, which rely in stead on inexpensive radars & cameras (Toyota, Chrysler) or two cameras (Subaru)