new

scripts/gyms/Sprint.py

@@ -282,7 +282,6 @@ class sprint(gym.Env):
         Draw.clear_all()
         self.player.terminate()
 
-
     def generate_random_target(self, position):
         while True:
             angle = np.random.uniform(0, 2 * np.pi)
@@ -312,7 +311,7 @@ class sprint(gym.Env):
             self.walk_distance = np.linalg.norm(self.walk_target - r.loc_head_position[:2])
         self.walk_rel_orientation = M.vector_angle(self.walk_rel_target) * 0.3
         # exponential moving average
-        self.act = 0.8 * self.act + 0.2 * action * action_mult * 0.7
+        self.act = 0.6 * self.act + 0.4 * action
 
         # execute Step behavior to extract the target positions of each leg (we will override these targets)
         lfy, lfz, rfy, rfz = self.step_generator.get_target_positions(self.step_counter == 0, self.STEP_DUR,
@@ -363,9 +362,9 @@ class Train(Train_Base):
     def train(self, args):
 
         # --------------------------------------- Learning parameters
-        n_envs = min(12, os.cpu_count())
+        n_envs = min(10, os.cpu_count())
         n_steps_per_env = 1024  # RolloutBuffer is of size (n_steps_per_env * n_envs)
-        minibatch_size = 64  # should be a factor of (n_steps_per_env * n_envs)
+        minibatch_size = 256  # should be a factor of (n_steps_per_env * n_envs)
         total_steps = 50000000
         learning_rate = 3e-4
         folder_name = f'Sprint_R{self.robot_type}'

scripts/gyms/sprint_demo.py

@@ -322,7 +322,9 @@ class sprint(gym.Env):
         self.sync()
         self.step_counter += 1
         obs = self.observe()
-        reward = r.loc_torso_velocity[0] - r.loc_torso_velocity[1] * 0.2
+        direction_error = abs(self.walk_rel_orientation)
+        direction_error = min(direction_error, 10)
+        reward = (r.loc_torso_velocity[0] - r.loc_torso_velocity[1] * 0.2) * (1 - direction_error / 10)
         if self.player.behavior.is_ready("Get_Up"):
             self.terminal = True