fix(PPO_Bot): SIGSEGV crash fixes + static buffers for thread safety

- bullets: seq[InFlightBullet] → array[4, InFlightBullet] + bulletCount
  (eliminates cross-thread heap realloc under ORC)
- hasFired: edge-triggered (cleared after state build, not level-triggered)
- round_counter parseInt: wrapped for empty/torn file → 0
- Static SVG + intent buffers to kill cross-thread heap realloc
- Tick-local alive/bulletData also fixed arrays

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-08-20 14:23:08 +02:00
parent 75e32e3315
commit 6ad51148f4
11 changed files with 239 additions and 52 deletions
+68 -3
View File
@@ -1,7 +1,7 @@
## PPO_Bot — enemy tracker + state vector wired into the game loop.
## Training: trajectory collected per tick, PPO update in background thread.
import std/[os, strformat, strutils, math, times]
import std/[os, strformat, strutils, math, times, algorithm]
import arraymancer
import tankroyale_botapi
import network
@@ -35,6 +35,7 @@ var
# Wire logStd tunable params into network module vars (read before initActorCritic)
logStdFloor = getEnvFloat("PPOB_LOG_STD_FLOOR", -3.0'f32)
logStdCeiling = getEnvFloat("PPOB_LOG_STD_CEILING", 0.5'f32)
initialLogStd = getEnvFloat("PPOB_INITIAL_LOG_STD", 0.0'f32)
# ── Structured log output ─────────────────────────────────────────────────────
@@ -59,11 +60,17 @@ proc hyperparmSnapshot(): string =
&"\"entropyCoeff\":{hpEntropyCoeff},\"valueLossCoeff\":{hpValueLossCoeff}," &
&"\"maxGradNorm\":{hpMaxGradNorm},\"gamma\":{hpGamma},\"lam\":{hpLam}," &
&"\"epochs\":{hpEpochs},\"miniBatchSize\":{hpMiniBatchSize}," &
&"\"logStdFloor\":{logStdFloor},\"initialLogStd\":{initialLogStd}"
&"\"logStdFloor\":{logStdFloor},\"logStdCeiling\":{logStdCeiling},\"initialLogStd\":{initialLogStd}"
const botJsonPath = currentSourcePath().parentDir / "PPO_Bot.json"
const weightsRoot = currentSourcePath().parentDir / "weights"
type
InFlightBullet = object
x, y: float64 # current position
vx, vy: float64 # velocity (pixels/tick)
power: float64
type PPOBot = ref object of Bot
tracker: EnemyTracker
buffer: TrajectoryBuffer
@@ -77,6 +84,12 @@ type PPOBot = ref object of Bot
lastActions: BotActions # previous tick's decoded actions (for state vector)
roundRewardSum: float32 # cumulative reward this round (for live display)
roundTicks: int # ticks this round
# Fixed-size bullet buffer — NO heap on the shared bot object. A seq here is
# allocated by the per-round bot thread and freed by the next round's thread
# (bot.bullets = @[] on round start) → foreign-heap free under --threads:on +
# ORC → SIGSEGV. N=4: the state vector only consumes the closest 3 slots.
bullets: array[4, InFlightBullet]
bulletCount: int
var ac = initActorCritic()
var gAdamStates: ACAdamStates # persists across rounds
@@ -98,6 +111,7 @@ method onRoundStarted*(bot: PPOBot, e: RoundStartedEvent) =
bot.hasLastTrans = false
bot.roundRewardSum = 0.0'f32
bot.roundTicks = 0
bot.bulletCount = 0
method onRoundEnded*(bot: PPOBot, e: RoundEndedEventForBot) =
inc roundCounter
@@ -184,6 +198,53 @@ method run(bot: PPOBot) =
while isRunning():
bot.tracker.deadReckon()
# Spawn bullet when enemy fired last scan tick. Edge-triggered: the flag is
# consumed later (after the state build) so a radar gap (deadReckon ticks)
# can't spawn k phantom bullets from one shot — exactly one bullet, and
# state index 12 still pulses 1 on the detection tick.
if bot.tracker.hasContact and bot.tracker.current.hasFired:
if bot.bulletCount < 4:
let power = bot.tracker.current.lastFirePower
let speed = 20.0 - 3.0 * power
# Approximate gun direction: bearing from enemy toward our position
let myX = getX(); let myY = getY()
let ang = arctan2(myY - bot.tracker.current.y, myX - bot.tracker.current.x)
bot.bullets[bot.bulletCount] = InFlightBullet(
x: bot.tracker.current.x,
y: bot.tracker.current.y,
vx: speed * cos(ang),
vy: speed * sin(ang),
power: power,
)
inc bot.bulletCount
# Advance in-flight bullets and prune those off-arena (in-place compaction
# into the fixed buffer — no per-tick heap churn).
let aW = float64(getArenaWidth()); let aH = float64(getArenaHeight())
var n = 0
for i in 0 ..< bot.bulletCount:
let b = bot.bullets[i]
let nx = b.x + b.vx; let ny = b.y + b.vy
if nx >= 0.0 and nx <= aW and ny >= 0.0 and ny <= aH:
bot.bullets[n] = InFlightBullet(x: nx, y: ny, vx: b.vx, vy: b.vy, power: b.power)
inc n
bot.bulletCount = n
# Convert to BulletData for state vector (closest 3 by distance to us)
let myX2 = getX(); let myY2 = getY()
var bulletData: array[4, BulletData]
var bulletDataCount = 0
for i in 0 ..< bot.bulletCount:
let b = bot.bullets[i]
bulletData[bulletDataCount] = BulletData(x: b.x, y: b.y, power: b.power)
inc bulletDataCount
# sort ascending by distance so the nearest threats fill slots 0-2
if bulletDataCount > 1:
bulletData.toOpenArray(0, bulletDataCount - 1).sort(proc(a, b: BulletData): int =
let da = hypot(a.x - myX2, a.y - myY2)
let db = hypot(b.x - myX2, b.y - myY2)
cmp(da, db))
setRadarTurnRate(bot.tracker.getRadarTurnRate(getX(), getY(), getDirection(), getRadarDirection()))
let botData = BotStateData(
@@ -203,7 +264,11 @@ method run(bot: PPOBot) =
let remainingGunAngle = abs(normalizeRelativeAngle(
directionTo(botData.x, botData.y, bot.lastActions.aimToX, bot.lastActions.aimToY) -
botData.gunDirection))
let state = buildStateVector(botData, bot.tracker, remainingGotoDistance, remainingGunAngle)
let state = buildStateVector(botData, bot.tracker, remainingGotoDistance, remainingGunAngle, bulletData, bulletDataCount)
# Consume the fired flag AFTER the state build: state index 12 saw the
# detection-tick pulse, and the next iteration's spawn check sees false —
# one shot → exactly one bullet, even across deadReckon gaps.
bot.tracker.current.hasFired = false
let (rawActs, logP) = ac.actorForward(state)
let value = ac.criticForward(state)
let ex = if bot.tracker.hasContact: bot.tracker.current.x else: botData.arenaWidth / 2.0