§ NAVIGATION
This is the synthesis behavior: seek a goal and avoid obstacles at the same time. The fly homes on a goal beacon — the chemotaxis odor cue, reused — and detours around a wall it senses with two bilateral feelers. The detour direction isn't hard-coded; it falls out of the feeler left−right asymmetry, arbitrated against the homing drive. It's the most robot-demo-compelling of the four — and the most honest: unlike escape, navigation has no clean real-circuit seam.
Two drives, arbitrated into a detour
A top-down view of the fly with the goal beacon ahead and a wall in the path. Two things are bilateral here: where the goal is (the odor L − R antenna difference, reused from chemotaxis) and where the wall is (two L vs R feeler fans reading obstacle proximity). Each yields a drive — a seek vector toward the goal and an avoid vector away from the wall — and the controller must arbitrate the two when they disagree. The resolved detour heading is just their sum. Hover, tap, or focus any element to read its role.
two drives, arbitrated
seek · toward goal
from the odor L − R antenna asymmetry — the homing drive. Alone, it walks into the wall.
avoid · away from wall
wall reads stronger on the left (proxL − proxR = +0.31), so the push is to the right — emergent.
seek ⊕ avoid→ detour right around the wall
L − R) and avoid the wall (feeler L − R) — and the detour is just their sum. The dodge direction is emergent, the same bilateral-asymmetry trick as chemotaxis and escape. Note: the feelers are a robotics rangefinder abstraction — real flies avoid obstacles with vision / optic flow, so this is a capability demo, deliberately not a neuron map.The geometry is analytic — the feeler readings are real ray casts against the wall and the vectors are their true sums, not artwork. The detour direction is emergent: nothing in the diagram hard-codes which way to dodge — it falls out of proxL − proxR, the same bilateral-asymmetry motif as chemotaxis (turn toward odor) and escape (bolt away from a loom).
No live demo here — on purpose
The arbitration above is analytic and real— true ray casts against the wall, true vector sums — but it's a frozen schematic, not a trained fly walking. A live place-the-goal / drag-the-obstacles demo, the kind perturbation and chemotaxis ship, would need something that doesn't exist yet: a navigation controller actually trained with the feeler inputs and exported alongside recorded rollouts and top-down trajectories — a new data-n compute bundle, the same shape as the data-c2 and data-ch bundles those live demos replay.
Rather than ship a stub or fake a controller, navigation stays what it honestly is: the exploratory, least connectome-aligned behavior of the four. Unlike escape, which drops onto a real LC4 / LPLC2 → Giant-Fiber circuit, the feelers are a robotics abstraction with no neuron behind them. It earns its place as the seek-plus-avoid synthesis, shown analytically — and the live trained demo is a flagged compute follow-up, not a promise rendered as a placeholder.
What it senses
A goal bearing and two obstacle feelers. The goal is the chemotaxis odor gradient, reused — bilateral, so the left−right antenna difference says which way it lies, exactly as in chemotaxis. The new sense is a pair of short-range feelers: obstacle proximity over the left forward field vs the right, written into the grid each control step the way every other sense is wired in the Sensing tab. So both questions — “where's the goal” and “where's the wall” — arrive as the same kind of left−right comparison.
The reward
Reach the goal, don't hit anything. Fitness rewards reaching the goal beacon and penalizes collisions (and rewards reaching it quickly). No term says which wayto go around a wall — the detour has to be discovered from the feeler asymmetry. As with the other behaviors, this is a task-specific objective: don't read its fitness scalar against walking, chemotaxis, or escape.
The result
The detour direction emerges from the feeler L−R asymmetry, arbitrated against the homing drive. It's warm-started from the chemotaxis forager, so with the feelers off it is the pure forager — it walks straight at the goal and into the wall. The avoidance is the added skill: when a wall reads nearer on one side, the controller turns the other way and detours around it, then re-homes — with nothing hard-coding “dodge away from the closer side.” The two drives are genuinely competing (seek pulls toward the goal, avoid pushes off the wall), and the behavior is the learned arbitrationbetween them. The visual above is that arbitration, frozen in one frame; a live place-the-goal demo would need a trained navigation controller we haven't run — it's a flagged compute follow-up, deliberately not faked here.
Connectome link
The honest one: there is no clean real-circuit seam here. This is the opposite of escape, which maps onto a real, mapped circuit (LC4 / LPLC2 → the Giant Fiber / DNp01) — the place a FlyWire sub-circuit literally drops in. Navigation has no such bridge. Real flies avoid obstacles with vision and optic flow, not short-range rangefinders; our two feelers are a robotics abstraction, a hand-built front-end with no neuron behind it. That makes navigation the most robot-demo-compelling behavior and the least connectome-aligned— a capability / synthesis demo, not a connectome bridge. Keeping that distinction visible is the point: the site stays calibrated about what's grounded in real biology (escape's seam) and what's a useful demo (these feelers). The real-biology direction to chase later is vision-based avoidance, not a feeler circuit — a different and harder Embodied rung.
The math behind this — equations and constants — is in the appendix.