[Dprglist] self driving trucks and the desert

[Paul Bouchier]

Didn't know David's email also went to dprg list, so I'm posting part of an
offline conversation I had with David in response to his email, because I
think good points are made all the way round. Read from the bottom up.

(I haven't figured out how to get gmail to play nicely with dprglist - tips
welcomed.)

===================
>From dpa

Hi Paul,

Thanks for the reply.  I think that perhaps part of the reality is that I
have developed a subsumption based reactive navigation method that I
believe is fairly unique --- the only other people I've encountered that do
the same have learned it from me.

As a consequence, people use the phrases "behavior base" and "reactive" to
describe what in their minds are simple wall followers and so forth.   They
don't envision a navigating subsumptive robot because they have never
encountered such and can't themselves conceive of it.  It took me quite a
while to realize that I have actually, as has been told to me, taken
subsumption to a "different level."

So I've found that there are folks all over the globe that have adopted my
methods, just not in the DPRG.   A  prophet in his own land... ;)

cheers!

David

===================

>From Paul

On 10/24/21 1:15 PM, Paul Bouchier wrote:
> > */[EXTERNAL SENDER]/* > > Hi David - thank you for your email. I always
appreciate your > thoughtful discussion, and you challenge me to think
critically. My > responses below are my invention of reasoning for the
choices the > system architects made - they never explained to me the
rationale for > the decisions they made. > > You make good points
throughout your email. However, what really > triggered my "desert driving"
analogy was observing the end pose of > the robot after avoiding the
obstacle - it was cocked at 20 degrees > to the straight line path between
start and finish. And similarly, > when it turned the corner into the long
hallway it swung wide (since > it didn't stop) then proceeded to hug the
wall that was closer to the > end goal. Those would not be considered good
road-driving behaviors - > highway vehicles should as far as possible
remain centered in their > lane, aligned with the lane direction, and
should return quickly to > that pose after an obstacle avoidance maneuver.
They should negotiate > bends or corners at a speed appropriate to the
curvature of the > bend. > > At a higher level, I think good and safe
driving in lanes benefits > from some amount of look-ahead. A global plan
provides the > look-ahead, and the local planner reacts to local obstacles.
If > Chris' office wasn't so filled with obstacles, we should have seen >
the local planner direct his robot back to the globally planned path, >
from where it could continue to execute the global plan. A global > plan
for highway driving should keep a vehicle centered in a lane. As > a safety
feature, the drive-by-wire subsystem of the trucks consumed > the
waypoints, and would decelerate appropriately to bring the > vehicle to a
halt at the end of a series of waypoints. This not only > provided smooth
stopping at the end of a mission, but if the path > generator crashed (e.g.
computer died), the drive-by-wire subsystem > would smoothly bring the
vehicle to a halt at the end of whatever > length of path it had been
given, no matter the curvature or > irregularity of the road. Thus a
shorter local path means the vehicle > would drive less distance without
obstacle sensors and avoidance > planning. Also, the presence of many
waypoints around a curve means > the drive-by-wire subsystem can calculate
an appropriate speed to > avoid roll-over while cornering, and slow down
prior to arriving at > the turn. > > Subsumption (in the simple
implementation used by robots) by contrast > is purely reactive, so doesn't
have the look-ahead that would keep it > cornering correctly if
higher-levels failed, or keep it on a good > global plan, well centered in
a lane. You can hypothesize all kinds > of fixes, maybe more waypoints, but
in the end I think you'd end up > creating some kind of look-ahead. You can
argue that the hallway is > like a lane, but the robot doesn't attempt to
do "good lane driving" > in it, and would require much more intelligence
and perhaps global > maps to do so. IOW your subsumption implementation
doesn't need > hundreds of waypoints because it doesn't try to drive well
in lanes, > or protect against failure of higher-level systems. The trucks
could > probably have used fewer, subject to the "driving without obstacle
> avoidance" tradeoff in case of failure, but waypoints are cheap, and >
dropping them every meter or so is easy. > > Maybe your point is that
"limited to driving in a desert" is too > constraining an analogy, because
the robot clearly can drive in a > lane, albeit not in a way that you would
call "good driving" if > performed on a highway. Karim put it best when he
said waypoint paths > attract the vehicle to them, and thereby keep it on a
global path > which has hopefully been planned well, vs. pure reactive
steering > driven by obstacle avoidance. > > There was actually a "road
following" mode of the army trucks, in > which they used lidar and camera
data to know where a road went, and > generated a path down the road. This
provided all the advantages of > look-ahead, while not requiring a global
plan - it just followed the > road - the road was the global plan. (It
didn't work very well, but > that's a whole different story.) > > In
convoying mode, the army trucks used all the sensors to try to > keep a
speed-dependent distance between the follower and its leader: > GPS from
both vehicles, lidar & radar ranging, and UWB ranging. It > also used those
sensors to generate hte follower path - it didn't > just blindly follow
waypoints dropped by the leader - rather, the > leader's waypoints were
part of what fed into follower path planning. > This made follower path
generation more robust in the face of radio > dropouts, GPS jumps, etc. > >
Thanks for providing motivation to think about how that system > worked > >
Paul
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dprg.org/pipermail/dprglist-dprg.org/attachments/20211025/f1642d98/attachment.html>