Behaviour trees

A behaviour tree is an authored priority list, read left to right and top to bottom. A composite walks its children by a rule; a task is a leaf that does something; a decorator is a condition attached to a node; a service runs on an interval for as long as a branch is active.

It compiles to a BehaviorTreeTemplate — a flat array of nodes in depth-first pre-order, immutable and shared by every agent running it — and each agent gets a BehaviorTreeInstance, which is a byte block and an active node.

⚠ It is not walked every frame. The tree keeps the active node and does nothing at all when nothing has changed: the active task ticks, active services tick if their interval has elapsed, and pending aborts are serviced. An agent walking across a courtyard costs one MoveTo.Tick and a service every 0.4 s.

Anything a designer must be able to read and predict: boss phases, scripted encounters, a guard's patrol-notice-chase-give-up loop. A tree is the planner you choose when "why did it do that" has to be answerable by pointing at a branch.

It is bad at the cross-product. Twenty conditions is a tree nobody can hold in their head, and a reactive condition costs a decorator on every branch it can interrupt. When the action set is open-ended and tuning by curve beats tuning by structure, a utility set is the better shape; when the interesting thing is a sequence nobody authored, GOAP is.

Register the actions, build a layout, author the tree, compile it, and give an agent its index.

Two kinds would have been enough to express every tree. Four is what makes one readable: "chase the player, checking every half second whether he is still visible, and give up if he stops being" is nine nodes in four levels with two kinds, and one task with a decorator and a service with four.

⚠ Decorator order on a node is significant. They evaluate top to bottom and the first failure stops the rest, so putting the cheap test above the expensive trace is your decision and the tree honours it. On the way out they unwind innermost-first, so an inverter under a force-success reads the way it is drawn.

⚠ Selector resumes and Priority does not. "Does a selector resume" is the question every implementation answers differently and silently, so they are two composites. A Priority re-walks from child zero every step, which is how a higher-priority branch takes over with no observer anywhere.

⚠ Parallel runs one main task, not N branches. Child 0 must be a task; child 1 is the branch that runs alongside it. True N-way parallelism makes the abort scope undefinable — two branches whose decorators want to abort each other — and the compiler refuses anything else.

A decorator declares ObserverAborts. When it is not None, a write to a key it reads re-tests it, and if the answer changed:

• Self — it is failing and its own subtree is running, so tear that down.
• LowerPriority — it is passing and something after its subtree is running, so take over.
• Both — both.

The test is two integer comparisons against the decorated node's pre-order range, which is what the flat layout exists for.

⚠ An abort takes effect at the start of the next step, not immediately. A task writes its own results during its tick — that is the ordinary case — and tearing it down from inside that write would destroy the state of the thing currently executing. The one-step latency is the price and it is stated rather than hidden.

⚠ A decorator reaches the siblings under its own parent composite, and no further. Unreal's abort reaches further up the tree, which is more powerful and is the subject of most of the confusion in its forums. The narrower rule is what lets the editor draw what a decorator can interrupt. The cost is real and worth knowing: a condition that becomes true deep inside a branch the agent has already walked past does not pull it back, because that branch is not running and nothing there is listening. Put the decorator on the branch you want interrupted, not inside it.

⚠ A decorator that declares an abort mode but reads no key can never fire. The compiler refuses it rather than letting it ship, because the symptom is "the AI sometimes gets stuck".

A time limit expiring, a cooldown ending, a target walking out of a cone: no key changed, so no observer can see it. Those decorators set Continuous and are re-tested every step while their branch runs — and when one goes false it fails its branch, so the composite moves on to its next child. That is different from an observer abort, which restarts the composite from child zero.

BehaviorTree.Subtree is spliced at compile time: the child's nodes are written into the parent's array in place of the calling node. Pre-order still equals priority across the boundary, so a decorator in the parent can abort a branch inside the child and the range test still means something. A cycle is refused by name.

RunSubtreeDynamicTask names its tree from a blackboard key, so it cannot be spliced and gets an instance of its own. The parent can abort the whole of it and can see nothing inside it — the honest cost of choosing a tree at run time.

A condition of your own. Everything per-agent goes in the span, because one decorator object is shared by every agent running the tree:

Reading the world into a key on a schedule — a service is a local sensor with an interval, which is why there is one implementation and two front ends onto it:

Stepping a tree by hand, which is what a headless test does:

BehaviorTreeInstance.Trace remembers what every node last returned, so the editor's canvas can tint the tree by it — active, path, succeeded, failed. An aborted node records as failed, because "why did the thing I was watching stop" is the question the tinting exists to answer.

⚠ It is off, and the array is allocated the first time it is turned on and never again. A per-node status byte is a per-agent cost paid to serve a panel nobody has open, and it is deliberately not in the memory block — the block is sized at load for every agent in the game and this is wanted by one of them at a time. See the AI debugger.

• Authoring a behaviour tree — the .vxbt file and the editor over it.
• Agents and actions — what a task is, and the system that steps a tree.
• The blackboard — what a decorator reads and what an abort observes.