claude solved dGPU power management in linux on the intel retina macbooks, so docking works properly
we're capitalizing on the pain of the math nerds and implementing a CFD system that can bake aerodynamic models and let us give the fly proprioception of airflow over its jet's wings
the pilots are in the fpga, real hardware! it's on the pcie bus, and the fruitfly pilots are flying both the hero and the enemies (NES and hypermapper still emulated)
after an overnight training run, training larvae in batches, here's our hero fly on the left and enemy flies on the right, dogfighting.
The model running the emulator is still the training version that runs on the 4090, running the sim thru the NES emu via our hypermapper dynamic rom
Next we'll hook up the actual FPGA thru the emulator
the fly still hasn't solo'd in the jet, but our hypermapper supports many enemies, and our fpga fly inference engine supports batchsize 16 so every enemy gets a fly brain too
banger
The baseline pilot is working through the same observation/action interface a future fly will use.
18/18 navigation cases passed, zero crashes.
Neutral controls: 3/18 passed, six crashes.
All 23 regression tests pass.
Clean capture: 900 frames verified pixel-for-pixel.
The demo shows left-bank, diving-bank, and high-climb recovery, with deliberate episode resets at 10 and 20 seconds.
This is conventional feedback control in our arcade model; no fly connected yet.
















