Skip to content

Release Notes 5.1.1

Jump to bottom
Dudiebug edited this page Apr 30, 2026 · 2 revisions

Release Notes - 5.1.1

This is the historical release note for the 5.1.1 combat reliability and first movement-network release. For the current 5.2.x movement brain-bank behavior, see Movement Brain Bank, AI Training, and Configuration.

Combat Reliability

Vanilla Attack Ordering Fix

Bot.attack() calls getBukkitEntity().attack(entity) before punch/swing animations. Previously, swing/punch could reset the attack charge before the damage calculation, causing ghost hits where the animation plays but no damage lands.

Charge-Aware Planning

The CombatDirector checks BotCombatTiming.chargeReady() before committing to attack branches. Bots wait for full attack charge before swinging, preventing low-damage spam and swing-block cycling.

Mace Recharge Planning

The mace smash branch accounts for custom gravity and required airtime. The Director does not commit to a mace smash unless the bot can realistically complete the jump-fall cycle and reach smash-ready charge.

Mace Airborne Tracking

When mid-mace-fall, tracking uses:

  • Ground clip tolerance to avoid premature landing detection.
  • Velocity-aware horizontal damping to stay centered on moving targets without overshooting.

Sweep Instrumentation

Sweep-check and sweep-skip telemetry were added to the combat trace system.

Combat Telemetry

Debug fields include critPred, sweepPred, chargeAtVanillaAttack, chargeAfterVanillaAttack, targetHp, and targetHpDelta. Enable with /bot combatdebug.

Movement Neural Network Overhaul

Movement-Only Network

The movement-controller neural network handles footwork: strafing, spacing, sprint/jump timing, approach/retreat, facing adjustment, and hold-position cooperation. It does not control combat. CombatDirector retains full authority over weapon selection, attack timing, and execution.

CombatIntent / MovementState Coupling

  • CombatDirector.plan() produces a CombatIntent.
  • The NN receives CombatIntent as part of its input.
  • After movement, MovementOutputApplier updates MovementState.
  • CombatDirector.execute() reads MovementState to validate timing windows before attacking.

Network Architecture

  • 37-value input in current 5.2.x builds.
  • 8-value output: forward, strafe, jump, sprint, retreat, facing, urgency, hold.
  • Configurable hidden layers.
  • tanh activation.

In-JVM GA Training

  • Tournament selection, uniform crossover, Gaussian mutation, and elite preservation.
  • Configurable population.
  • All training runs in the JVM; no Python, ONNX, GPU, or external tooling.

Loadout Mixing

The first version used a weighted loadout pool. In 5.2.x this became the movement brain-bank loadout mix under ai.training.loadout-mix and ai.training.loadout-mixes, with mixed rounds training eligible specialist families.

Brain Persistence

The first version used a single brain.json. In 5.2.x this became the manifest plus per-family bank under ai/movement/. Legacy compatible ai/brain.json files import as general_fallback.

Commands Introduced

  • /ai brain <status|load|save|reset> manages movement brains.
  • /ai movement <amount> <name> spawns bots using the saved movement bank.
  • /ai reinforcement ... [movement-controller|legacy] chooses training mode. In 5.2.x, movement-controller is the default when mode is omitted.
  • /bot loadoutmix <mix> distributes loadouts across live spawned bots.
  • /bot combatdebug <name|all> <on|off> toggles combat trace logging.

New Loadouts

Added: vanilla, axe, smp, pot, spear.

See Loadouts for details.

Configuration Notes

The 5.2.x config uses:

  • ai.movement for movement-bank runtime settings.
  • ai.movement.bank for manifest/per-brain persistence settings.
  • ai.training.max-round-minutes for the reinforcement round cap.
  • ai.training.loadout-mix and ai.training.loadout-mixes for weighted training loadout selection.
  • ai.training.curriculum-family for focused specialist training.

See Configuration for the current full reference.

Compatibility

  • Current 5.2.x release jars target Paper 26.1.2 only.
  • Java 25.
  • Brain files from earlier versions may be incompatible. Use /ai brain reset to start fresh if validation fails.

Known Cautions

  • Training spawns many bots fighting simultaneously. Monitor server TPS and reduce population if needed.
  • Changing movement layer shape after training invalidates saved brains. Reset and retrain.
  • /bot create bots do not use the movement bank. Use /ai movement to spawn bots with the trained movement bank.

Clone this wiki locally