Upload train_pixelcopter.py

train_pixelcopter.py

@@ -0,0 +1,88 @@
+import gymnasium as gym
+from gymnasium import spaces
+import numpy as np
+from stable_baselines3 import PPO
+
+# -------------------------
+# Side-Scrolling PixelCopter Environment (Medium/Certification Friendly)
+# -------------------------
+class PixelCopterCertEnv(gym.Env):
+    def __init__(self, screen_width=50, screen_height=10, gap_size=6):
+        super().__init__()
+        self.screen_width = screen_width
+        self.screen_height = screen_height
+        self.copter_y = self.screen_height // 2
+        self.copter_velocity = 0
+        self.gravity = 0.25
+        self.lift = -0.9
+        self.done = False
+        self.timestep = 0
+        self.max_timesteps = 500
+        self.gap_size = gap_size
+
+        self.wall_gap_positions = [np.random.randint(1, self.screen_height - self.gap_size -1) 
+                                   for _ in range(screen_width)]
+
+        self.action_space = spaces.Discrete(2)
+        self.observation_space = spaces.Box(
+            low=0, high=self.screen_height, shape=(self.screen_width + 1,), dtype=np.float32
+        )
+
+    def reset(self, seed=None, options=None):
+        self.copter_y = self.screen_height // 2
+        self.copter_velocity = 0
+        self.done = False
+        self.timestep = 0
+        self.wall_gap_positions = [np.random.randint(1, self.screen_height - self.gap_size -1) 
+                                   for _ in range(self.screen_width)]
+        obs = np.array([self.copter_y] + self.wall_gap_positions, dtype=np.float32)
+        return obs, {}
+
+    def step(self, action):
+        # Apply action
+        if action == 1:
+            self.copter_velocity += self.lift
+        self.copter_velocity += self.gravity
+        self.copter_y += self.copter_velocity
+        self.copter_y = np.clip(self.copter_y, 0, self.screen_height)
+
+        # Move walls left
+        self.wall_gap_positions = self.wall_gap_positions[1:]
+        last_gap = self.wall_gap_positions[-1]
+        new_gap = last_gap + np.random.choice([-1,0,1])
+        new_gap = np.clip(new_gap, 1, self.screen_height - self.gap_size -1)
+        self.wall_gap_positions.append(new_gap)
+
+        # Check collision with first wall
+        gap_top = self.wall_gap_positions[0]
+        gap_bottom = gap_top + self.gap_size
+        if self.copter_y <= gap_top or self.copter_y >= gap_bottom:
+            self.done = True
+            reward = -5
+        else:
+            reward = 1
+
+        self.timestep += 1
+        if self.timestep >= self.max_timesteps:
+            self.done = True
+
+        obs = np.array([self.copter_y] + self.wall_gap_positions, dtype=np.float32)
+        return obs, reward, self.done, False, {}
+
+# -------------------------
+# Training
+# -------------------------
+env = PixelCopterCertEnv(screen_width=80, screen_height=10, gap_size=6)
+model = PPO("MlpPolicy", env, verbose=1)
+
+print("Training started...")
+model.learn(total_timesteps=500_000)  # Enough for certification
+print("Training finished!")
+
+model.save("ppo_pixelcopter_cert")
+print("Model saved as 'ppo_pixelcopter_cert.zip'")
+
+
+
+
+