jalf/TIR_PROJ
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Sheep training flock of 10 fix?

Browse Source
This commit is contained in:
Johnny Fernandes
2026-04-23 23:20:23 +01:00
parent a13f5d0ff0
commit f68dea44da
1 changed files with 16 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
training/herding_env.py
+16 -12
View File
@@ -53,9 +53,9 @@ class HerdingEnv(gym.Env):
# ----------------------------------------------------------------------- # -----------------------------------------------------------------------
# Reward weights (progress-based potential shaping + sparse bonuses) # Reward weights (progress-based potential shaping + sparse bonuses)
# ----------------------------------------------------------------------- # -----------------------------------------------------------------------
W_DRIVE = 2.0 # flock COM moved toward pen (per metre, per step) W_DRIVE = 2.0 # progress: flock COM moved toward pen
W_COLLECT = 1.0 # flock radius shrank (per metre, per step) W_COLLECT = 0.5 # progress: flock radius shrank
W_APPROACH = 0.3 # stable position signal: dog close to flock COM W_ALIGN = 0.5 # position: dog on anti-pen side of flock COM
W_PEN_BONUS = 5.0 # per sheep penned W_PEN_BONUS = 5.0 # per sheep penned
W_COMPLETE = 20.0 # all sheep penned W_COMPLETE = 20.0 # all sheep penned
W_STEP_COST = 0.002 # time penalty W_STEP_COST = 0.002 # time penalty
@@ -288,23 +288,27 @@ class HerdingEnv(gym.Env):
com, radius, _ = self._flock_stats() com, radius, _ = self._flock_stats()
com_dist = float(np.linalg.norm(com - self.PEN_CENTER)) com_dist = float(np.linalg.norm(com - self.PEN_CENTER))
# Progress rewards: positive when flock moves toward pen or compacts # Progress rewards: positive when state improves
drive_progress = (self._prev_com_dist - com_dist) * self.W_DRIVE drive_progress = (self._prev_com_dist - com_dist) * self.W_DRIVE
collect_progress = (self._prev_radius - radius) * self.W_COLLECT collect_progress = (self._prev_radius - radius) * self.W_COLLECT
self._prev_com_dist = com_dist self._prev_com_dist = com_dist
self._prev_radius = radius self._prev_radius = radius
# Approach: stable position signal so the dog has a gradient toward # Alignment: reward dog for being on the anti-pen side of the flock
# the flock even when the sheep are not actively fleeing # COM, gated by proximity so only nearby positioning counts.
active_mask = ~self.penned[:self.n_sheep] # +1 = dog directly behind flock, -1 = dog on pen side (wrong).
if active_mask.any(): d_dog_com = float(np.linalg.norm(self.dog_pos - com))
dog_to_com = float(np.linalg.norm(self.dog_pos - com)) if d_dog_com > 0.1 and com_dist > 0.1:
approach = -(dog_to_com / (2 * self.FIELD)) * self.W_APPROACH pen_dir = (self.PEN_CENTER - com) / com_dist # COM → pen
dog_dir = (self.dog_pos - com) / d_dog_com # COM → dog
cosine = -float(np.dot(pen_dir, dog_dir)) # +1 when opposite
proximity = max(0.0, 1.0 - d_dog_com / self.FLEE_DIST)
alignment = cosine * proximity * self.W_ALIGN
else: else:
approach = 0.0 alignment = 0.0
reward = drive_progress + collect_progress + approach reward = drive_progress + collect_progress + alignment
reward += newly_penned * self.W_PEN_BONUS reward += newly_penned * self.W_PEN_BONUS
reward -= self.W_STEP_COST reward -= self.W_STEP_COST
if n_penned == self.n_sheep: if n_penned == self.n_sheep: