Neural Shepherd

Behind the game

Neural Shepherd is a co-op puzzle disguised as a quick field job. Ten sheep are nervous, the pen waits on the right, and one dog is trying to work with you rather than obey you. You control the shepherd's pressure: stand too close and the flock scatters, stand too far away and it drifts without direction. The dog can make a useful push, but it cannot repair every stray you allow to escape. A good sixty-second run has a rhythm of gathering, guiding, and letting the group move before asking for the next correction.

Begin by making the flock compact. Instead of chasing the sheep that looks most lost, approach from the outside so your presence nudges several animals toward the same middle. Once there is a group, leave a clear lane toward the pen and avoid standing in front of it. The right moment to move is often just after the dog has gone around the far side: your body can keep the near side from leaking while its pressure sends the mass forward. If one sheep breaks away, do not drag the entire flock after it without thought. Regroup the nearby animals first, then collect the stray when the cost is smaller. The score to care about is the pen count, so controlled progress beats a frantic chase.

The dog uses a neural policy learned from rewards for moving sheep home. It was not given a script that says “circle left” or “push this exact animal.” From its egocentric view it learned useful tendencies, such as swinging wide behind a nearby cluster and applying pressure toward the pen. Because it sees a local part of the flock, your role matters: keep the sheep close enough that its choice can affect more than one of them. Watch for its wide arc and complement it rather than competing for the same side. The little pasture becomes a readable partnership between two different kinds of attention. You see the whole scene; the dog acts on the pressure it can feel. When both actions agree, a loose, skittish crowd suddenly flows like one animal toward home.