Things that decide for themselves

orblit_agent has two layers. They often get mixed together, and they are better kept apart.

Steering answers where to go: a force, worked out from the world, that moves something this frame. Seek, flee, arrive, wander, separate, align, cohere, pursue, evade.

Behaviour trees answer what to want: which of those to run, and when to stop.

Take an agent that runs away when it is frightened and wanders the rest of the time. That is a behaviour tree picking between two steering behaviours. Build it as one thing and you get a state machine, and you have to rewrite it every time you add a state.

Steering behaviours add up

Behaviours return forces, and forces add. That is the whole model. It is why a flock is separate + align + cohere with weights, and not a flocking algorithm.

Two things the textbooks get wrong

separate must not normalise. The point of it is that it pushes harder the closer things get. Normalise the result and you throw that away and get an even shove. It looks right until things bunch up, and then it doesn’t.

arrive does overshoot. Most write-ups say it doesn’t. It does. What you get from it is a controlled approach speed, which is the thing you wanted anyway. Plan for the overshoot instead of assuming it away.

Behaviour trees

Sequences, selectors, decorators and leaves: the standard vocabulary. A tick returns running, succeeded or failed, and a node that was running on the last tick is resumed rather than restarted.

Unlike the sampled parts of the engine, a behaviour tree really does step. A decision made last tick is meant to stick, and that is history rather than a value at a time. See sampled, not stepped for where the line falls.

Spatial queries

Steering needs to know about neighbours, and asking every agent about every other agent is quadratic. orblit_collide gives you the spatial hash for it.

The XOR hash trap

A spatial hash shaped like x*a ^ y*b ^ z*c collides on symmetric coordinates, and the symptom is a strange one. A shape lands in the same bucket twice and then pairs with itself, so an agent tries to move away from where it already is. Mix the coordinates one after another instead of with XOR.