Jack Crossfire's blog View Details
Archive for May, 2008
Posted by Jack Crossfire | May 30, 2008 @ 02:16 PM | 6,608 Views
U can throw away your barometers, pitot tubes, washout filters, neural networks, sonar & Kalman filters. Just buy a uBlox 5.

Turned off the neural network and got solid hovering in the wind using straight PID feedback. Very fast response to wind gusts. There's no lag on any axis & it's dead accurate.

They're either using carrier phase smoothing or throwing a genetic algorithm at the CA codes. It really was just a matter of waiting for it to get cheaper.

All the missions have been navigation model 6. Altitude in ground effect is still very unstable. Above ground effect it's the most stable we've had. Still has vortex ring state & ballooning in the right conditions.

Sadly, we've probably seen the end of neural networks for navigation & 590 of our 600 Mhz are now wasted. There's still a neural network for cyclic, but with this level of navigation it's not necessary. Aside from autonomous takeoff & landing, the uBlox didn't leave many problems to solve.

VicaCopter under uBlox5 control (2 min 40 sec)
...Continue Reading
Posted by Jack Crossfire | May 29, 2008 @ 07:26 PM | 6,723 Views
uBlox5 is a monster. The latest positioning engines R way ahead of what we had a year ago in terms of latency & velocity readings. May have been a lucky satellite placement, however.

The barometer seems to be a dud. In the wind it detected pressure waves from which no meaningful climb rate could be extracted. Not a problem since the uBlox can beat it.

uBlox 5 is not based on an Atmel chip. Based on the stagnation of the Antaris line, they'll be handing out extra mortgage bailouts in San Jose today....Continue Reading
Posted by Jack Crossfire | May 29, 2008 @ 03:17 AM | 6,180 Views
The uBlox gods were dissapointed with our hand soldered Altoids creation, so they sent down a one of a kind module. PCB & integrated antenna, almost as compact as the EM406. Commute 1: figure out it's set to 38400 baud UBX protocol & not the default.

Neural washed up

The neural washout filter didn't go very well. The problem seems to be this is a recurrent neural network & back propagation doesn't work very well for a recurrent neural network. Time to go back to genetics. We do have the steady state filter processing barometer & GPS in all their unsynchronized glory.
Posted by Jack Crossfire | May 26, 2008 @ 09:26 PM | 6,612 Views
In a daylight expedition to the new test range, it didn't feel very close to the dumpy apartments but it felt a lot closer to the houses than on The Goog. Renters R one thing. Home owers R a different beast. Being $1 million underwater on a mortgage + copter noise = future Baghdad suicide bomber.

To integrate the barometer in the original neural network, would have to modularize the initialization code & have 3 modules, each handling 1-3 axes from an independant navigation source. Not planning on ever doing this.

The neural washout filter is priority 1. The neural washout filter will be trained on the ground. Want to combine all 3 axes in a single module. Even then, it's a complete waste of the 600Mhz CPU. Not an easy move. Maybe that CPU can train a neural PID replacement....Continue Reading
Posted by Jack Crossfire | May 25, 2008 @ 06:47 PM | 6,718 Views
Don't try it unless you've got tons of mortgage bailouts on the way, you really need to prove your manliness, or another EOS 5D is just 2 weeks of rent. Sewing thread works. It has to be clean & dry of course. It won't make the sensor spotless, but it can remove large spots that air alone can't. In exchange, U get a sensor full of lint which air can remove.

Unfortunately we seem to have a tiny chemical spot on the sensor....Continue Reading
Posted by Jack Crossfire | May 25, 2008 @ 03:23 AM | 6,813 Views
With gas now over $4.20, thinking more about ways to use the golf course instead of the park. Anything would have to be silent & invisible.

What we need is a flight box on wheels suitable for rough terrain. It would store a lead acid battery & laptop.

After reviewing GATech's T-Rex videos, have concluded the flight dynamics model probably has more mileage as an alternative to carrier phase GPS. Now playing with 2 ideas but no plan. 1 is to apply the barometer to the existing model.

The other is again shrinking the adaptive part by making more assumptions in the fixed part. The next frontier is the adaptive washout filter.

With the washout filter, angle of attack goes in & velocity change comes out, letting U model short periods of navigation without GPS. It needs a bandwidth & gain which change depending on payload & weather.

The neural network should have automatically developed a washout filter, but it didn't because we solved it the wrong way. To save clockcycles, it didn't take its own output as an input. Training it the right way would require filtering a bunch of past data points for 1 back propagation.

If it uses 30 sec of past data for each back propagation, that's 30*25=750 solutions. If it takes 7500 back propagations to learn how to fly, that's 5,625,000 solutions. If the Gumstix only does 15,000 solutions/sec, that's 6 minutes of flying to learn how to fly.

Definitely could not train a washout filter in flight.

Have our levitation portfolio.

Magnetic levitator compilation (3 min 47 sec)

Posted by Jack Crossfire | May 24, 2008 @ 04:48 AM | 6,691 Views
With 2% interest & tons of government aid, houses in San Francisco R now $10 million. That's enough for 20 suborbital flights counting inflation. The space traveler becomes famous & gets rich on book royalties & promotions. The home ower lives in the united version of Antarctica.

Meanwhile, have decided humans are programmed to obsess & brood, endlessly, over anything that hovers. Maybe some prehistoric ancestor millions of years ago was saved by something that hovered & every male organism since has kept the program to worship hovers.
Posted by Jack Crossfire | May 23, 2008 @ 03:52 AM | 6,672 Views
Magnet UAV Rides again (0 min 44 sec)

That's all we had planned for Magnet UAV. The rest is just getting the heater to heat the solution & maybe force some oscillation. Then it'll sit for a month.
Posted by Jack Crossfire | May 22, 2008 @ 03:18 AM | 6,665 Views
The solution to the PIC reset requirement is an analog circuit to pulse it. Analog computers R great fun. The reset pulse could have been generated by 555 timers of equal complexity, but this reuses our op-amp.
Posted by Jack Crossfire | May 21, 2008 @ 03:24 AM | 6,485 Views
Remembered the magnet UAV crashes any computer near it. Would make a good E bomb if it had any range. Also remembered the PIC needs a manual reset after powering up to work properly.

As for VicaCopter, adding pressure climb rate to the neural network is a buster. Creating training tables where the 2Hz barometer, 1Hz GPS, & 25Hz IMU R aligned is a buster.
Posted by Jack Crossfire | May 20, 2008 @ 03:18 AM | 6,266 Views
Originally wanted VicaCopter to be a motion control platform for video. Clearly, that isn't going to happen without major investments in GPS & a huge airframe. VicaCopter still has motion control possibilities with the still camera & software stabilization. Making it happen would require a video downlink to set waypoints based on camera view.

VicaCopter has possibilities in just attitude hold mode for video. Never tried that one because flying in attitude hold is difficult in wind.

The lack of convenient flying room makes ground based motion control the most likely use of VicaCopter's spare parts. Since the high end consumer video cameras R mostly used for panning around inanimate objects, a motion controlled still camera could be a big win.

Isn't it a good time to bring back...the magnet UAV.
Posted by Jack Crossfire | May 18, 2008 @ 02:33 AM | 6,413 Views
After 1.5 years of happy test flights, finally got kicked off the golf course. The days of easy access to airspace were over. Had previously flown no closer than 50ft to the dumpy apartments & that became 150 ft after discovering the enhanced noise from carbon blades & sideways orientation. Not sure it ever made more noise than an air conditioner.

Have since heard stories of photographers getting arrested because their super quiet point & shoot camera made too much noise during a classical concert which they had permission to photograph.

Like classical music listeners paying $90 a ticket, the average Calif*ahan pays $50,000/year to live here & is probably going to spend $100 million over their lifetime. The tiniest noise at all which isn't an air conditioner, an SUV, a TV, or something normal can drive them crazy.

Renters were probably complaining about noise forever. Condo owers feared crashes into their $1 million "property". One of them finally followed us home & took down the address. Wonder what else they noted, as regards expensive belongings & times the door is unlocked.

The budget couldn't have sustained the previous rate of flying anyways. It doesn't matter how good U R. With the T-Rex beyond a certain point, X amount of flight hours cost Y amount of bucks. Have never had an orientation loss with the T-Rex. It's all equipment nowadays.

Ironically, used the golf course specifically because there were never any people on it. The search is on for another test range. Definitely need to make this operation more portable. Flying is definitely going to be rare.

In the radio department, have left the radio system on, the 3.3V regulator under full load, in 80' heat, for 1 hour without any failure. Took the radio out of range & it still recovered from the highest packet corruption. Wiggled the connectors & transmitter ribbon cable, ran the transmitter battery to exhaustion & didn't have any problems.
Posted by Jack Crossfire | May 16, 2008 @ 01:12 PM | 6,457 Views
Repairing the peg board had another advantage, like finding out what caused the crash. Found an incomplete flight recording on the flash. The Dean connector did not glitch. Obviously, it lost the radio & went into failsafe mode. That held it in a stable hover for 3 sec before killing the engine. It impacted before the filesystem could sync.

The PIC didn't fail, because she would have spun out of control before engine shutdown. The PIC has to constantly trim rudder & cyclic to damp rates. Have never seen the XBee get anywhere close to losing 3sec of data.

That leaves a glitch in the radio connectors or a 3.3V regulator
overheating. Never saw any voltage anomaly in the flight recording. Need to leave the radio system & all the 3.3V parts on for hours, test range, & wiggle connectors until the problem repeats. Only then do we install a heat sink.
Posted by Jack Crossfire | May 15, 2008 @ 01:09 AM | 6,633 Views
So after ringing up the $60 in replacement carbon, the guy said Dean connectors R the most reliable. Well, clearly another board rework is in order & out go our prefab plans. The simplest solution is a redundant lead from the balancing tap & replacement of the main Molex connector with a Dean.

Reverse current protection with the non standardized balancing taps would require a custom adaptor for each battery. U couldn't do it without raising GND by 0.5V.

Note these glitches R not with the Molex because motor studdering was captured in vicacopter_a560_full.mov.
Posted by Jack Crossfire | May 14, 2008 @ 04:20 AM | 6,096 Views
So had our first complete loss of power during a 200 ft hover. Appears to be a glitch in a Dean connector. Dean connectors glitched many times in the past on the ground. The only solution is either different connectors or redundant Dean connectors.

The crash was a global killer. Nothing survived, not even bacteria. Most of the carbon fiber parts were destroyed. The main board was destroyed. Micromag3 was damaged.

Undecided on whether to replace it with a professionally made surface mount board. Going to be a long time before the next flight.

Had 15 more minutes of light wind tests before the crash. Problems with blade chopping vibration were solved & we were down to fine tuning pressure altitude. Position oscillation could be reduced by rotating magnetic declination. Cyclic & direction of motion could be matched. Got more blue sky in the background....Continue Reading
Posted by Jack Crossfire | May 13, 2008 @ 03:54 AM | 5,992 Views
U can save a few billion dollars & forget about carbon blades. Put the plastic on again & the roll anomalies went away for the most part. Need a few more hours to confirm that one.

The problem is when blades injest air from tail rotor exhaust. The carbon blades make a loud snare drum sound. Plastic blades make a quieter tympani sound. The theory is plastic flexes in the turbulance, transmitting less high frequency vibration from the blade tips to the accelerometer. The louder the noise, the more vibration the accelerometer experiences.

Wrote a VRS algorithm. Jam the collective if climb rate descends below a minimum. The trick is getting it to solve VRS without defeating intended descents.
Posted by Jack Crossfire | May 12, 2008 @ 02:56 AM | 6,267 Views
Results have been pretty terrible with the carbon blades. They seem noisier than plastic & transmit more vibration. Carbons chopping their own air causes more noise & position oscillation than plastic.

Vibration modes show up in accelerometer bias. Everything is stable, then suddenly a roll to port. Got good lighting, however. Had remarkably stable air on the hill.

Found a bug where altitude error was being used to derive horizontal velocity.

Another bug in CHDK causes it to take only 1 shot every 10 seconds using the shutter wire. The workaround is to take several shots using the shutter button & then use the wire....Continue Reading
Posted by Jack Crossfire | May 11, 2008 @ 03:10 AM | 6,059 Views
Tried some 20mph wind. Facing sideways into this wind, 90' rotations of accelerometer derived attitude were common & she often flew away. When not impacted by the wind on the wrong side, she could hover. Also, somehow the camera only shot every 10 seconds & it didn't show the true exposure in viewfinder mode.
Posted by Jack Crossfire | May 10, 2008 @ 03:51 AM | 6,138 Views
Put carbon blades back on & began testing the 2Hz barometer mode. Carbon blades have less coning, less wind influence, less drag, & cost 3x more than plastic. The reduced drag shows up in the yaw trim.

Ran out the transmitter battery in a 1 hour test flight marathon. Basically lived in the air.

We now have a 0.0001 gyro bias weighting. Had 2 anomalies where she entered catastrophic position oscillations in a strong vibration mode of the blades. Reminiscent of a stuck cyclic. Nothing unusual in any flight recording & couldn't reproduce it anymore.

Noticed how noisy she is from the sides despite being very quiet from the front....Continue Reading