fix: exp14 finetune load warm-start model without temp env to prevent second spawned car

agent/experiments/exp14_finetune_v5.py

@@ -158,39 +158,31 @@ def log(s):
 phase_defs = [ (PH1_STEPS, 0.4), (PH2_STEPS, 0.2) ]
 
 # create initial env and model (warm start)
-# Load model with base action space (throttle_min=0.2). We'll enforce a runtime
+# IMPORTANT: load the model WITHOUT an env, then attach exactly one env.
-# throttle FLOOR during phase 1 via a wrapper, but keep the action space unchanged.
+# Creating a temporary env just for loading opens a second TCP connection and
-loaded_env = VecTransposeImage(DummyVecEnv([make_env_base(0.2, throttle_floor=None)]))
+# spawns a second car in the sim (right lane + left lane issue).
 if os.path.exists(WARM_PATH):
-    log(f'Loading warm-start model from {WARM_PATH} using base throttle_min=0.2 env')
+    log(f'Loading warm-start model from {WARM_PATH} without creating a temp env')
-    model = PPO.load(WARM_PATH, env=loaded_env, device='cpu')
+    model = PPO.load(WARM_PATH, device='cpu')
     # override lr and schedules — ensure lr_schedule callable exists
     model.learning_rate = LR
     try:
         model.lr_schedule = get_schedule_fn(LR)
     except Exception:
         model.lr_schedule = None
-    # update optimizer param groups to new LR
     try:
         for pg in model.policy.optimizer.param_groups:
             pg['lr'] = LR
     except Exception:
         pass
-    # Create the training env using base action space but enforce throttle_floor at runtime
+    # Create exactly one training env and attach it
     first_throttle_floor = phase_defs[0][1]
     env0 = VecTransposeImage(DummyVecEnv([make_env_base(0.2, throttle_floor=first_throttle_floor)]))
     model.set_env(env0)
-    # Close the loaded_env used only for model loading to avoid leaving a stale
-    # TCP connection (which would create an extra vehicle in the simulator).
-    try:
-        loaded_env.close()
-    except Exception:
-        pass
 else:
     log('No warm-start found — creating fresh model with base throttle_min=0.2')
     env0 = VecTransposeImage(DummyVecEnv([make_env_base(0.2, throttle_floor=phase_defs[0][1])]))
     model = PPO('CnnPolicy', env0, learning_rate=LR, verbose=1, device='cpu')
-    loaded_env.close()
 
 steps_done = 0
 best_reward = float('-inf')