Dropped direct LPLC2/LC4 readouts, expanded simulation with downstream neurons and additional LC-family neurons. Remade readout net. The drone is steered exclusively with descending motor neurons which natively fire in response to visual flow injected into fly's optical train.

Working on correct LC11 and LPLC2 simulation proved to be challenging due to lack of correct inhibition. Trying to expand brain simulation even further with neuron dendrite position data to achieve natural optical filtering with inhibitory inputs.

How does the fly brain process visuals so much faster than Astra computer use which takes a few seconds per action/reason?

maybe try it on a real drone? it actually looks promising ngl with custom asic chips made specificly to run fly brain inference this could be quite cheap in large batch size

What simulator are you using?

Can you use those neurons that use continuous learning to reward hack him to do deliveries ?