Movement Network

Movement Network

The movement-controller network handles footwork only. It never selects weapons, changes hotbar slots, attacks, blocks, uses items, places crystals or anchors, or executes utility combat actions. CombatDirector owns combat; movement brains own locomotion.

For the full bank architecture, persistence layout, training mixes, evaluation commands, and regression guard, see Movement Brain Bank.

Tick Flow

1. CombatDirector.plan(...) creates a typed CombatIntent.
2. MovementBrainRouter selects a bank brain by committed lock, then CombatIntent.branchFamily, then range/role fallback, then general_fallback.
3. MovementInput.from(...) creates the normalized observation vector.
4. MovementNetwork.evaluate(...) produces movement preferences.
5. MovementOutputApplier.tryApply(...) applies walking, jumping, sprinting, retreating, strafing, and facing adjustment only.
6. MovementState is reported back to CombatDirector.execute(...).

CombatIntent Fields

Field	Purpose
branchFamily	Semantic MovementBranchFamily selected by CombatDirector
playId	Optional scanner/play telemetry
desiredRange / desiredRangeBand	Desired spacing and coarse range band
rangeUrgency	How urgent the spacing correction is
lockFamily, lockReason, lockUntilTick, interruptible	Commit/latch ownership
plannedAction	Debug/action label consumed by the stored plan
wantsCritSetup, wantsSprintHit, wantsHoldPosition	Movement hints for legal combat timing
isCommitted, commitProgress, weaponRange	Current phase and range metadata

Input Schema

The current observation schema has 37 values:

• relative position and distance
• bot and target velocity
• facing dot/cross
• grounded/falling/sprinting/recent-jump flags
• reachability/obstruction flags
• desired range and urgency
• crit/sprint/hold intent flags
• commit state and weapon range
• one-hot branchFamily.* fields for every MovementBranchFamily

The observation schema hash is exported in the brain manifest and evaluation reports. Changing fields requires a schema/hash change and retraining or reset.

Output Schema

The network outputs 8 movement values:

Output	Effect
forwardPressure	Move toward or away from target
strafePressure	Left/right pressure
jumpDesire	Jump when over threshold and grounded
sprintDesire	Sprint when over threshold and not retreating
retreatDesire	Prefer backing out
facingAdjustment	Small yaw adjustment
urgency	Movement speed multiplier
holdPosition	Cooperate with Director hold requests

Fallbacks

Missing, incompatible, or invalid specialist brains fall back to general_fallback. Missing or corrupt files are reported by /ai brain status and by /ai evaluate exports. Non-network bots spawned through /bot create continue using legacy movement.