29 lines
1.0 KiB
Plaintext
29 lines
1.0 KiB
Plaintext
first trial was with plain gap-error minimization reward,
|
|
|
|
# --- FIX 2: Smoother Reward Function ---
|
|
# Reward function
|
|
reward = 1.0 # Survival bonus
|
|
|
|
# Distance Penalty (Squared is smoother than linear for fine control)
|
|
reward -= (gap_error * 100)**2
|
|
|
|
# Orientation Penalties
|
|
reward -= (roll_angle * 10)**2
|
|
reward -= (pitch_angle * 10)**2
|
|
|
|
next, added the following:
|
|
|
|
contact_points = p.getContactPoints(bodyA=self.podId, bodyB=self.trackId)
|
|
has_contact = len(contact_points) > 0
|
|
|
|
# Don't terminate on contact.
|
|
# Instead, penalize it, but allow the episode to continue so it can try to fix it.
|
|
# if has_contact:
|
|
# 5.0 is painful, but surviving 100 steps of pain is better than immediate death (-50)
|
|
reward -= len(contact_points)
|
|
|
|
# at this point, we still either stick or fall, no hovering training has been achieved.
|
|
|
|
# Tried increasing lambda value and starting at optimal all the time.
|
|
|
|
#Tried reducing entropy and resetting all params but allowing for full range of motion without bolts - 7 pm ish |