So the MRF49XA we've battled finally got an upgrade. It's the MRF89XA &
it has a 32x bigger FIFO, hardware packetizing, hardware CRC, & now
requires a 12.8Mhz crystal so throw away all your 10Mhz crystals.
Work slowly resumes on neural networks for Marcy 1. This time, just settling for
feedforward prediction of navigation points. Have no confidence in it.
Once again, the problem with any completely artificial intelligence
feedback controller is computing the long term offset.
Now that we're optimizing CPU usage for neural networks again,
discovered if the sonar period is too long, it causes the audio
processing to exceed the cache & slow it way down.
The Aiptek P-HD got its 1st ground video of a difficult takeoff to
400ft & landing. Landing on the narrow slit between the weeds still
takes manual control. Extremely high wind caused pretty crazy navigation
errors on the takeoff.
Kept making demo videos in case someone ever read our resume. Standard
autonomous operations with VicaCopter in HD & very windy conditions.
Autonomous takeoffs, ascents, landings, hovers. Marcy maneuver didn't
work very well on that 400ft descent because of the wind.
The problem with wind is the high angle of attack causes GPS to lose
satellites. She dropped to 7 satellites at 400ft.
Finally, have some more sky writing. The star trails of Scorpio are deliberately visible in the background. It was always conspicuous in all our years of flying. Used to think it was a sickle marking death to anyone who ventured to the center of the galaxy.
The smile inducing technology really killed us in this shot.
So many diydroners are unemployed, it's time to start a group.
Comments
From what I read on Jack's vicacopter page, the primary motivation for using the NN was to get away from having to adjust PID values as conditions (or the platform) changed.
A little over a year ago, I squeezed the lwneuralnet (the one that Jack used on vicacopter) onto an atmega128 (with 32K RAM) but almost nothing else was running on it and I think I only got the update rate to 10Hz (but then again, everything was using floating point)
http://lwneuralnet.sourceforge.net/
I didn't carry it any further than that I'm afraid because I had much larger problems (and still do!).
Good Luck
Earl