ConnectomePilot

Wires the MaleCNS connectivity graph into a simulated mobile robot: sensor input drives neural activity and a policy readout produces corrections on top of a base navigation controller. The recipe that worked best trains internal connections and readout jointly for 65,536 steps, then freezes the connections and continues on the readout to 4,194,304. On 500 held-out maps with three seeds the mean success rate is 85.1% at about a million steps and 88.6% at four million, with 89.4% for the model picked by the pre-declared rule. Longer training did not help monotonically.

Official site ↗

#robot#navigation#residual-control#ppo#staged-training