Date: Sat, 11 Feb 2006 16:25:04 -0600 (CST)
Subject: lessons of stanley
X-UID: 134

The last week was my "15 minutes of Warhol fame" for this project. While 
gratifying, it was also distracting. However, I think I needed a week off 
to figure out what to do next now that hardware development is done.

There was a presentation at the PARC Forum about "Stanley's Brain":

http://www.parc.com/cms/get_article.php?id=503

The presentation is:

http://www.parc.com/events/forum/media/v1119.mp3
mms://216.93.180.194/parc_forum/v1119.wmv

Note that you can save the WMV stream using:

mencoder -oac mp3lame -ovc lavc mms://216.93.180.194/parc_forum/v1119.wmv

I find that the MP3 is probably the best as the WMV is only 160x120 video 
anyway. The entire presentation is quite long, somewhere around 90 
minutes. This talk is a goldmine of information.

I'm still analyzing the presentation but here's what I've taken out of it 
so far.

1. Hand-tuned probabilistic heuristics are required with manual driving
    capability for collecting live sensor data - closed form analytical
    methods like Kalman filters are not enough

2. Reliability matters - it is better to have a small number of reliable
    systems even if they perform worse than you'd like

3. Development is iterative - test drive robot in field, analyze data,
    recode software, and repeat this over and over

The Stanford effort had enormous resources. I didn't fully realize how 
much until hearing this talk. It involved multiple departments and private 
companies. Volkswagen's drive-by-wire vehicle was a one-of-a-kind 
prototype. They had a "dream team" of people and resources behind the 
machine.

I just picked up a USB joystick and inverter. With the stick, I'll be able 
to manually drive the robot. This is necessary to collect sensor data for 
filter design. The presentation basically confirms the impracticality of 
designing a control system with an analysis mindset, even if you have lots 
of money. You have to collect experimental data from field operation and 
then respond to that.

With the inverter, I can run the laptop in the car for as long as I still 
have gasoline in the tank. I can also recharge the NiMH battery packs for 
the drive motors. Those will drain quickly while the digital SLA battery 
should last for one to two hours.

The first priority now is a remote control system. I know that some people 
are thinking, "Duh! That's obvious!" But the purpose of the remote control 
is to collect live run sensor datasets for design and tuning of the 
control system. It's not an extra option but required, sort of like how 
the motor control electronics could not be tested until the 
microcontroller was working. Some subsystems are critical points in the 
entire system. Without them, you can't connect them together and run 
end-to-end or even know if things are broken.

I'm going to try and get a system together by tomorrow morning so I can, 
weather permitting, collect the first dataset. I'm probably going to use a 
flash drive key mounted on the robot to record the sensor data, especially 
video frames. Driving will be done over WiFi for now. 900 MHz radio will 
be used just as in the DARPA race for three functions: remote kill, pause, 
restart.