flyOnlineLearning

Continual reward search over the actual deployed GPU synapses. Eight-second windows reward pickup and previously unvisited 0.5 m odometry cells. Repeated distance earns no reward; revisits, stopped motion and bumper contact are penalized. Unproductive closed turns receive an additional loop penalty. Odometry integrates incremental ground-motion measurements; the learner receives no trash coordinates, room map or route.

Sparse proposals remain within a trust region around the calibrated checkpoint (±8 tonic-bias conductance units, ±5 other weights), with bounded exploration. Unproductive proposals cannot become the accepted policy. This prevents the unbounded reverse-motor drift that previously produced endless spirals.

The learner changes weights only. Native sensory spikes traverse the added trained synapses and native motor spikes remain the sole actuator source. Online trials occur in changing surroundings, so acceptance is not a paired causal performance comparison. regression.luau verifies repeated-circle penalties, stationary penalties, productive rewards and weight bounds over 98 circular trials.