Thread Tools
This thread is privately moderated by Jack Crossfire, who may elect to delete unwanted replies.
Sep 01, 2009, 12:26 PM
Registered User
Jack Crossfire's Avatar
Thread OP
Discussion

the last flight ... OF CROSSFIRE


Alright kids. Wednesday we're doing a 60m aerial shoot of Mountain View & then we're done flying.

Flying UAV's on the golf course is just killing us. We're not recovering from Air Force Heroine. We're dying just on routine golf course photo shoots that used to be our vacation. We need a new gig. Enjoy the Marcy Foam, 3DOF IMU, VicaCopter, & spin copter. The blog will move to another time slot.





Now the last of the flying blog notes.

Marcy Foam vs. flag in the wind (1 min 56 sec)


Tough video to watch. Tough video to watch.

Exposed a lot of problems with flag payloads in the wind. Did U see how the tail rotor was so far over it shredded the radio cable? Did U see how horizontal velocity was maximized to keep the flag from pulling it away? Very psychotic. At least the Marcy Foam got improved performance.

spin copter closeout

Lots of questions arose on spin copter. Does the fuselage have too much momentum to slow down for half a revolution? Would the possible speed change give much horizontal velocity? What good is it if you're not also going to have a pushbroom camera?



spaceship blog post

Spaceships don't have accelerometers & magnetometers for sensing attitude. They only use gyros & photograph the stars to reset their IMU's. That makes VicaCopter essentially a spaceship.

MAV blog post

There is a plan to develop 2 systems for a future airframe:

Ground based navigation in polar coordinates:
2 sonar receivers & 1 IR transmitter on the ground
1 sonar transmitter & 2 IR receivers on the airframe
Sense altitude & radius by measuring sonar delay
Sense heading with IR receivers & gyro

Solenoid thrust vectoring:
mount tail motor on spring
Control with direct drive magnet.

problem with plywood fuselage & magnetometer

So basically, the plywood warped when it was installed & may change shape some more in flight. The result is the magnetometer & IMU aren't aligned & she goes into spirals. Spirals R caused by the compass reading deviating because the attitude estimate is wrong. What U need is a $150 tilt compensated compass from $parkfun. Fuggedaboutit.

USING LWNEURALNET RECURRENTLY

So with lwneuralnet, you call straight net_compute's for every step in time. Use a master network & unique inputs for every step. U need to store all the neuron outputs & network inputs for every net_compute step. Network inputs are the first layer outputs.

Then U call net_compute_output_error exactly once on the last iteration. This sets the errors on the last outputs.

Now net_train_batch to begin the training pass. This must be modified. backward_pass must start on the neuron outputs from the last time step & repeat working backward in time. Modify backward_pass to recall the neuron outputs for each time step, copy the first layer's errors in the current time step to the last layer's errors in the previous time step. backward_pass needs to store errors for all time steps instead of accumulating.

adjust_deltas_batch needs to go through all time steps & recall the errors & outputs for each time step. The deltas R properly accumulated in the master network.

Next, call adjust_weights_batch to train the master network. This averages the deltas over the number of time steps.


NOTE ABOUT LEAD COMPENSATION

The velocity term in your PID equations provides rate damping but not lead compensation. To get lead compensation, U need to add velocity to the position. The fully damped & lead compensated equation has a velocity feedback & position + velocity feedback.
Last edited by Jack Crossfire; Sep 01, 2009 at 01:20 PM.
Sign up now
to remove ads between posts
Sep 01, 2009, 11:06 PM
Registered User
I've really enjoyed your blog and learned allot from it. I have gotten lots inspiration from what your are able to accomplish without spending gobs of money.

I look forward to the next installment.

Reed

Quote:
Originally Posted by Jack Crossfire
Alright kids. Wednesday we're doing a 60m aerial shoot of Mountain View & then we're done flying.

Flying UAV's on the golf course is just killing us. We're not recovering from Air Force Heroine. We're dying just on routine golf course photo shoots that used to be our vacation. We need a new gig. Enjoy the Marcy Foam, 3DOF IMU, VicaCopter, & spin copter. The blog will move to another time slot.





Now the last of the flying blog notes.

http://www.youtube.com/watch?v=GXc-iHXv5o4

Tough video to watch. Tough video to watch.

Exposed a lot of problems with flag payloads in the wind. Did U see how the tail rotor was so far over it shredded the radio cable? Did U see how horizontal velocity was maximized to keep the flag from pulling it away? Very psychotic. At least the Marcy Foam got improved performance.

spin copter closeout

Lots of questions arose on spin copter. Does the fuselage have too much momentum to slow down for half a revolution? Would the possible speed change give much horizontal velocity? What good is it if you're not also going to have a pushbroom camera?



spaceship blog post

Spaceships don't have accelerometers & magnetometers for sensing attitude. They only use gyros & photograph the stars to reset their IMU's. That makes VicaCopter essentially a spaceship.

MAV blog post

There is a plan to develop 2 systems for a future airframe:

Ground based navigation in polar coordinates:
2 sonar receivers & 1 IR transmitter on the ground
1 sonar transmitter & 2 IR receivers on the airframe
Sense altitude & radius by measuring sonar delay
Sense heading with IR receivers & gyro

Solenoid thrust vectoring:
mount tail motor on spring
Control with direct drive magnet.

problem with plywood fuselage & magnetometer

So basically, the plywood warped when it was installed & may change shape some more in flight. The result is the magnetometer & IMU aren't aligned & she goes into spirals. Spirals R caused by the compass reading deviating because the attitude estimate is wrong. What U need is a $150 tilt compensated compass from $parkfun. Fuggedaboutit.

USING LWNEURALNET RECURRENTLY

So with lwneuralnet, you call straight net_compute's for every step in time. Use a master network & unique inputs for every step. U need to store all the neuron outputs & network inputs for every net_compute step. Network inputs are the first layer outputs.

Then U call net_compute_output_error exactly once on the last iteration. This sets the errors on the last outputs.

Now net_train_batch to begin the training pass. This must be modified. backward_pass must start on the neuron outputs from the last time step & repeat working backward in time. Modify backward_pass to recall the neuron outputs for each time step, copy the first layer's errors in the current time step to the last layer's errors in the previous time step. backward_pass needs to store errors for all time steps instead of accumulating.

adjust_deltas_batch needs to go through all time steps & recall the errors & outputs for each time step. The deltas R properly accumulated in the master network.

Next, call adjust_weights_batch to train the master network. This averages the deltas over the number of time steps.


NOTE ABOUT LEAD COMPENSATION

The velocity term in your PID equations provides rate damping but not lead compensation. To get lead compensation, U need to add velocity to the position. The fully damped & lead compensated equation has a velocity feedback & position + velocity feedback.


Quick Reply
Message:
Thread Tools

Similar Threads
Category Thread Thread Starter Forum Replies Last Post
First (and last) flight of the "AP Beast" boomerace Aerial Photography 11 Aug 09, 2003 06:01 PM
Last Flight of Bomber 31 SchiessCo Life, The Universe, and Politics 7 Jan 15, 2003 11:33 AM
Alas, Wingo's last flight Marksl Foamies (Kits) 7 Apr 26, 2002 09:16 PM
The last flight of my second lanier U-2 AMTJIM Foamies (Kits) 4 Jun 22, 2001 05:22 AM
Help determining altitude of last flight jdmartin8 Parkflyers 8 May 19, 2001 12:57 AM