new best

scripts/gyms/Sprint.py

@@ -285,12 +285,12 @@ class sprint(gym.Env):
     def generate_random_target(self, position):
         while True:
             angle = np.random.uniform(0, 2 * np.pi)
-            X = position[0] * 10 * np.cos(angle)
-            Y = position[1] * 10 * np.sin(angle)
-            if -13 <= X <= 13 and -8 <= Y <= 8:
+            x = position[0] + 10 * np.cos(angle)
+            y = position[1] + 10 * np.sin(angle)
+            if -13 <= x <= 13 and -8 <= y <= 8:
                 break
 
-        self.walk_target = np.array([X, Y])
+        self.walk_target = np.array([x, y])
 
     def step(self, action):
 
@@ -362,9 +362,9 @@ class Train(Train_Base):
     def train(self, args):
 
         # --------------------------------------- Learning parameters
-        n_envs = min(10, os.cpu_count())
+        n_envs = min(12, os.cpu_count())
         n_steps_per_env = 1024  # RolloutBuffer is of size (n_steps_per_env * n_envs)
-        minibatch_size = 256  # should be a factor of (n_steps_per_env * n_envs)
+        minibatch_size = 64  # 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

@@ -292,7 +292,7 @@ class sprint(gym.Env):
         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,