Finally got around to viewing some more quad copter videos & it's not looking good for the current, manually controlled system. There have been many startups, but Skydiddyo has blown it out of the park.  This is really a game changer in autonomy.  It does everything I tried to do with the ground vehicles for the last 5 years, manely following an athlete while directly ahead instead of behind, avoiding very dense obstacles, going at very high speed.  

While neural networks are required for any startup to get funded, they're probably only being used for recognizing the athlete with a chance of them being used to match features in the depth mapping.  The navigation is done with very high fidelity 3D mapping (SLAM), from 13 fisheye cameras instead of the 4 that other quad copters use.  The exact resolution & framerate wasn't given.  While it's a total failure as a $2500 consumer product, it's worth a certain amount of money to have it, with that level of miniaturization, in a developer kit that you can apply to any ground vehicle.  It's only a matter of time before it happens.

In the absence of a developer kit or money, the odroid might be enough to get some kind of lead mode without obstacle avoidance, with only 1 or 2 cameras.  20 years ago, they would have given away bits of their source code.  Those days are over.  Can only imagine they're taking advantage of better object tracking in the last 5 years to get a very accurate motion vector of the athlete.  The motion vector can be used to predict future position.

The running robot is not practical as a quad copter, because it needs to move cargo & last longer than 5 minutes.  The mane problem is keeping it on surfaces it can drive on.  Based on the demos, it would be quite difficult to stay on a 10ft wide path sometimes on the left or right of the athlete, sometimes with a curb, sometimes flat, sometimes with 45mph cars, while also avoiding obstacles.  This was where detecting the path became critical & all methods fell apart when they encountered shadows.  Perhaps very precise depth mapping can detect the path when the edges are high enough.  When the edges are too flat for depth mapping to detect, it could be safe enough to drive over them & just rely on athlete tracking.

Meanwhile, the Tamiya Lunchbox has proven a real turd since being converted to a lower CG with cargo space, so it might be worth spending the $300-$1000 for a better platform.  Another 7.4 mile drive sucked 3787mAh or 508 mAh/mile & ground off a lot of the tires.  The 7 minute pace & cargo proved a real buster in the current configuration.  The Futaba at 6V without servo saver constantly buzzes.  Spent all of it driving maybe 100 meters between tweeking a PID constant.

It's worth considering a feature where it takes ranges of values for the PID constants & automatically homes in on constants which give the least oscillation.  Unfortunately, in 1 year everything is going to have a Skydiddy level of autonomy.  A remote controlled pacer with no autonomy is going to be terribly old fashioned.