It's the National Ignition Facility, the most powerful laser in the world. $4 billion. 500 Terrawatts. Enough power to feed 413,223 time machines. It requires 5 hours between each 4ns laser firing to cool down.
It was too windy to get very high. Only got 1 shot at 400ft. Angle of attack got to the abort limit where GPS could be lost & conventional control laws don't work.
Next, it was back over Pillar point.
Can't go very high in fog.
Finally, flew up over Napa again & for the 1st time, didn't crash. The wind in this area was the highest she ever flew in. Saw pretty insane angles of attack. GPS reception was real lousy. The video was made from several flights with the camera at different angles. 3 fans of freedom on the wing made strange noises in the wind.
Noted 2 anomalies in the video: a rubber band came off the camera & radio dropouts
caused several pitch ups. Got 1 in the video. Aft propeller was scraping something when it hit yaw limits.
Looking over the lwneuralnet sourcecode again, the net_begin_batch, net_train_batch, & net_end_batch functions are basically the definition of backpropagation through time with obscure wording.
The neural network is recurrent but not being trained that way. Was always a bit stumped on how to train a recurrent network to optimize its own flying. They only know how to recall ideal answers. In the past, fed tables from PID equations.
1 way is to just randomly change weights & select better performing networks over long periods of time. Maybe have a 2nd network be an emprical model of the flight characteristics. Then there's taking the times when it hit the right position & training off those.
Well, the only way we see this commissioned turkey going autonomous now is ground based machine vision. Even then, it's not going to be stable enough to fit indoors. A ground camera can sense roll. The magnetometer can sense pitch. Sonar would still be needed for position.
Golf course grass on the UAV? It happens to the best of us, just like marriage, death, & taxes. Autonomy without attitude sensing isn't going so well again. Her neural network was overloading her CPU.
Comments
Just a small comment on the neural network you are using. Feed forward networks are excellent for prediction of future values, but you should consider the number of nodes in the hidden layer. Usually a large number of hidden nodes, i.e. more hidden nodes than input nodes, tend to learn the data set which you are training from. You should consider using around half the number of input nodes as hidden nodes. By doing this you will learn the characteristics of the data set, and not the data itself.
/Niklas
What kind of autopilot are you using? Neural network???