[FoRK] Autonomous vehicles

J. Andrew Rogers andrew at ceruleansystems.com
Fri Jan 16 12:01:13 PST 2009

On Jan 16, 2009, at 8:53 AM, Stephen Williams wrote:
> Specifically, the current DSLAM algorithms are a huge improvement,  
> and involve insight that can be used in other problems.  And they  
> are built on probabilistic relational reasoning and learning  
> (Bayesian / Markov with approximate reasoning algorithms for  
> efficiency).

The SLAM algorithms are getting better, but those are generally based  
on brittle algorithms like SIFT and limited by the relatively poor  
ability represent complex spatial models.  There are still some non- 
trivial theoretical computer science problems to solve for a robust,  
general-case solution.

What we have now makes a nice proof of concept, but they are  
unacceptably brittle and "dumb" for reasons that will not be solved  
with mere tweaking of the existing algorithms. Within certain limits  
they work well enough, but a lot of the autonomous vehicle navigation  
efforts are upgrading hot air balloons in an attempt to get to the  
moon, and it is predictable that we should expect diminishing returns  
unless the deeper limitations are addressed.  Otherwise, like fusion  
power plants, the technology will perpetually be just around the corner.

> We already have armed autonomous fliers, and that is an arena where  
> you could have some serious failures and vulnerabilities.

Autonomous fliers are a much simpler case that can safely ignore a  
number of hard problems that affect ground vehicles, allowing them to  
be dumber and still get the job done safely.  Cruise missiles with  
adaptive/reactive flight plans have been with us since (IIRC) the late  
1980s, and autopilot devices have been around forever.

>> Also, I would observe that the commercial infrastructure for such  
>> things that is being worked on includes deep and pervasive real- 
>> time tracking of individuals, but I will leave it to the civil  
>> libertarians as to whether or not that is a good thing.
> That is something that is, or could be, coming from several  
> directions.  It doesn't really need any autonomous vehicle AI.  If  
> you have a cell phone, or for that matter, a face in a camera- 
> infested area, you're already trackable.

Yeah, I was more talking about the case where the  large number  
(50-100) of sensors in your vehicle are tagged and back-hauled in near  
real-time.  When you consider the range of sensors that increasingly  
exist in modern cars, that builds a more detailed picture than I  
expect most people would be comfortable with.  This capability is  
being designed and built into some vehicles now -- obviously not self- 
driving cars -- as a component of the infrastructure for eventual self- 
driving vehicles, and there are already a few test bed areas.

This type of infrastructure is needed to fill some capability gaps for  
truly autonomous vehicles that having smart SLAM and GPS do not fill,  
but once you've grabbed the bits you need it is almost free to grab  
the rest of the sensor data.


J. Andrew Rogers

More information about the FoRK mailing list