new

scripts/gyms/sprint.py

@@ -56,6 +56,7 @@ class sprint(gym.Env):
 
         self.walk_rel_orientation = None
         self.walk_rel_target = None
+        self.walk_target = None
         self.walk_distance = None
         self.act = np.zeros(16, np.float32)  # memory variable
 
@@ -232,10 +233,12 @@ class sprint(gym.Env):
         w = self.player.world
         t = w.time_local_ms
         self.reset_time = t
-
-        distance = 15 - r.loc_head_position[0]
-        self.walk_rel_target = (15, self.Gen_player_pos[1])
+        self.generate_random_target()
+        distance = np.linalg.norm(self.walk_target[:2] - self.Gen_player_pos[:2])
         self.walk_distance = distance
+        self.walk_rel_target = M.rotate_2d_vec(
+            (self.walk_target[0] - r.loc_head_position[0], self.walk_target[1] - r.loc_head_position[1]),
+            -r.imu_torso_orientation)
         self.walk_rel_orientation = M.vector_angle(self.walk_rel_target)
 
         for _ in range(25):
@@ -279,40 +282,24 @@ class sprint(gym.Env):
         Draw.clear_all()
         self.player.terminate()
 
-    def change_target(self):
-        original_angle = self.walk_rel_orientation
-
-        orientations = random.choice([-1, 1])
-        random_angle_delta = orientations * np.random.uniform(10, 75)  # 单位是度
+    def generate_random_target(self, x_range=(-12, 12), y_range=(-9, 9)):
 
-        # 新的角度
-        new_angle = original_angle + np.radians(random_angle_delta)  # 转换为弧度
-
-        # 计算新的目标向量，保持原来的距离不变
-        new_walk_rel_target = np.array([
-            15,
-            np.sin(new_angle) * self.walk_distance * 3
-        ])
-
-        # 保存新的目标向量
-        self.walk_rel_target = new_walk_rel_target
+        x = np.random.uniform(x_range[0], x_range[1])
+        y = np.random.uniform(y_range[0], y_range[1])
 
+        self.walk_target = np.array([x, y])
     def step(self, action):
 
         r = (self.
              player.world.robot)
 
         w = self.player.world
-        current_time = time.time()
-        if current_time - self.last_target_update_time > 2:
-            self.change_target()
-            self.last_target_update_time = current_time
 
         internal_dist = np.linalg.norm(self.internal_target)
         action_mult = 1 if internal_dist > 0.2 else (0.7 / 0.2) * internal_dist + 0.3
         self.walk_rel_target = M.rotate_2d_vec(
-            (15 - r.loc_head_position[0], self.Gen_player_pos[1] - r.loc_head_position[1]), -r.imu_torso_orientation)
-        self.walk_distance = np.linalg.norm(self.walk_rel_target)
+            (self.walk_target[0] - r.loc_head_position[0], self.walk_target[1] - r.loc_head_position[1]), -r.imu_torso_orientation)
+        self.walk_distance = np.linalg.norm(self.walk_target[:2] - self.Gen_player_pos[:2])
         self.walk_rel_orientation = M.vector_angle(self.walk_rel_target) * 0.3
 
         # exponential moving average
@@ -349,17 +336,15 @@ class sprint(gym.Env):
         robot_speed = r.loc_torso_velocity[0]
         direction_error = abs(self.walk_rel_orientation)
         direction_error = min(direction_error, 10)
-        reward = robot_speed * (1 - direction_error / 10)
-        if np.linalg.norm(r.loc_torso_position[:2] - self.walk_rel_target,ord='fro') < 1:
-            reward += 50
+        reward = robot_speed * (3 - 0.5 * direction_error / 10)
+        if self.walk_distance < 1:
+            reward += 20
+            self.generate_random_target()
 
         if self.player.behavior.is_ready("Get_Up"):
             self.terminal = True
-        elif w.time_local_ms - self.reset_time > 40000:
-            self.terminal = True
-        elif r.loc_torso_position[0] > 14.5:
+        elif w.time_local_ms - self.reset_time > 100000:
             self.terminal = True
-            reward += 500
         else:
             self.terminal = False
         return obs, reward, self.terminal, {}