Fixed yaw_dif

Positive reward for stoping at final position.
Positive reward for correct orientation at final position.

carla_environment.py

@@ -246,27 +246,30 @@ class World:
                     carla.Location(x=-25.534311, y=54.460903, z=0.112781), \
                     carla.Rotation(pitch=0.000000, yaw=-175.922913, roll=-6.221135))
         
-        pos_diff = math.sqrt((target.location.x - x)**2 + (target.location.y - y)**2)
-
         v = math.sqrt((vx)**2 + (vy)**2)
+        pos_diff = math.sqrt((target.location.x - x)**2 + (target.location.y - y)**2)
+        target_yaw = (-175.922913) / 180
+        yaw_dif = math.sqrt((target_yaw - yaw)**2)
         
         r = -0.1
         r += -0.01 * pos_diff
         
-        LEARN_FINAL_ORIENTATION = False
+        LEARN_FINAL_ORIENTATION = True
         if LEARN_FINAL_ORIENTATION:
-            target_yaw = (-175.922913) / 180
-            yaw_dif = math.sqrt((target_yaw**2) - (yaw**2))
             r += -0.01 * yaw_dif
 
-        INSENTIVE_MOVING = False
+        INSENTIVE_MOVING = True
         if v < 0.01 and INSENTIVE_MOVING:
             r -= 0.02
 
         done = False
-        if pos_diff < 1.5:
+        if pos_diff < 1.2:
             done = True
             r += 100
+            if yaw_dif < 0.01:
+                r+= 100
+            if v < 0.01:
+                r+=50
             
         if self.collision_sensor.collision is not None:
             r -= 100

run_scripts/manual_carla.py

@@ -19,25 +19,61 @@ c.load_from = 'Carla_2_256__128__128_0.9995_0.001_1DBL'
 c.load_mem = True
 c.load_ann = True
 
+# o = copy.deepcopy(c)
+# o.name = 'JTAP_4'
+# o.force_cpu = True
+# o.render = True
+# o.learn = True
+# o.env_type = 'Carla'
+# o.net_layout = [256, 128, 128]
+# o.save_to = 'simple/'
+# o.load_from = 'Carla_JTAP_0_256__128__32_0.9995_0.001_1DBL'
+# o.load_mem = True
+# o.load_ann = False
+# o.learn_offline = True
+# o.offline_epochs = 75000
+# o.offline_batchsize = 256
+# o.learn_iterations = 1
+# o.offline_validate_every_x_iteration = -1
+# o.learn_online = False
+# o.eps_decay = 0.9995
+# o.learn_rate= 0.001
+# o.run_episodes = 15
+
 o = copy.deepcopy(c)
-o.name = 'JTAP_5'
+o.name = 'JTAP_0'
+o.force_cpu = True
 o.render = True
 o.learn = True
 o.env_type = 'Carla'
-o.net_layout = [128, 32, 32]
-o.save_to = 'test/'
-o.load_from = 'Carla_JTAP_4_128__32__32_0.9999_0.0005_1'
-o.load_mem = True
+o.net_layout = [256, 128]
+o.save_to = 'orientation/'
+o.load_from = ''
+o.load_mem = False
 o.load_ann = False
-o.learn_offline = True
-o.offline_epochs = 1500
-o.learn_iterations = 1
-o.offline_validate_every_x_iteration = -1
 o.learn_online = True
-o.eps_decay = 0.9999
-o.learn_rate= 0.0005
-o.run_episodes = 500
+o.eps_decay = 0.9995
+o.learn_rate= 0.001
+o.run_episodes = 750
 
+validate = copy.deepcopy(c)
+validate.name = 'JTAP_Validate'
+validate.render = True
+validate.learn = True
+validate.env_type = 'Carla'
+validate.net_layout = [256, 128]
+validate.save_to = 'simple/'
+validate.load_from = 'Carla_JTAP_0_256__128__32_0.9995_0.001_1DBL'
+validate.load_mem = False
+validate.load_ann = True
+validate.learn_offline = False
+validate.offline_epochs = 1500
+validate.learn_iterations = 1
+validate.offline_validate_every_x_iteration = -1
+validate.learn_online = True
+validate.eps_decay = 0.9995
+validate.learn_rate= 0.001
+validate.run_episodes = 10
 
 # t = copy.deepcopy(c)
 # t.render = True