diff --git a/alr_envs/alr/mujoco/reacher/alr_reacher.py b/alr_envs/alr/mujoco/reacher/alr_reacher.py
index 4477a66..b436fdd 100644
--- a/alr_envs/alr/mujoco/reacher/alr_reacher.py
+++ b/alr_envs/alr/mujoco/reacher/alr_reacher.py
@@ -39,11 +39,18 @@ class ALRReacherEnv(MujocoEnv, utils.EzPickle):
         reward_dist = 0.0
         angular_vel = 0.0
         reward_balance = 0.0
+        is_delayed = self.steps_before_reward > 0
+        reward_ctrl = - np.square(a).sum()
         if self._steps >= self.steps_before_reward:
             vec = self.get_body_com("fingertip") - self.get_body_com("target")
             reward_dist -= self.reward_weight * np.linalg.norm(vec)
-            angular_vel -= np.linalg.norm(self.sim.data.qvel.flat[:self.n_links])
-        reward_ctrl = - np.square(a).sum()
+            if is_delayed:
+                # avoid giving this penalty for normal step based case
+                # angular_vel -= 10 * np.linalg.norm(self.sim.data.qvel.flat[:self.n_links])
+                angular_vel -= 10 * np.square(self.sim.data.qvel.flat[:self.n_links]).sum()
+        if is_delayed:
+            # Higher control penalty for sparse reward per timestep
+            reward_ctrl *= 10
 
         if self.balance:
             reward_balance -= self.balance_weight * np.abs(
@@ -56,63 +63,66 @@ class ALRReacherEnv(MujocoEnv, utils.EzPickle):
         return ob, reward, done, dict(reward_dist=reward_dist, reward_ctrl=reward_ctrl,
                                       velocity=angular_vel, reward_balance=reward_balance,
                                       end_effector=self.get_body_com("fingertip").copy(),
-                                      goal=self.goal if hasattr(self, "goal") else None,
-                                      joint_pos = self.sim.data.qpos.flat[:self.n_links].copy(),
-                                      joint_vel = self.sim.data.qvel.flat[:self.n_links].copy())
+                                      goal=self.goal if hasattr(self, "goal") else None)
 
     def viewer_setup(self):
         self.viewer.cam.trackbodyid = 0
 
+    # def reset_model(self):
+    #     qpos = self.init_qpos
+    #     if not hasattr(self, "goal"):
+    #         self.goal = np.array([-0.25, 0.25])
+    #         # self.goal = self.init_qpos.copy()[:2] + 0.05
+    #     qpos[-2:] = self.goal
+    #     qvel = self.init_qvel
+    #     qvel[-2:] = 0
+    #     self.set_state(qpos, qvel)
+    #     self._steps = 0
+    #
+    #     return self._get_obs()
+
     def reset_model(self):
-        # qpos = self.np_random.uniform(low=-0.1, high=0.1, size=self.model.nq) + self.init_qpos
-        qpos = self.init_qpos
+        qpos = self.init_qpos.copy()
         while True:
             self.goal = self.np_random.uniform(low=-self.n_links / 10, high=self.n_links / 10, size=2)
+            # self.goal = self.np_random.uniform(low=0, high=self.n_links / 10, size=2)
+            # self.goal = np.random.uniform(low=[-self.n_links / 10, 0], high=[0, self.n_links / 10], size=2)
             if np.linalg.norm(self.goal) < self.n_links / 10:
                 break
         qpos[-2:] = self.goal
-        qvel = self.init_qvel# + self.np_random.uniform(low=-.005, high=.005, size=self.model.nv)
+        qvel = self.init_qvel.copy()
         qvel[-2:] = 0
         self.set_state(qpos, qvel)
         self._steps = 0
 
         return self._get_obs()
 
+    # def reset_model(self):
+    #     qpos = self.np_random.uniform(low=-0.1, high=0.1, size=self.model.nq) + self.init_qpos
+    #     while True:
+    #         self.goal = self.np_random.uniform(low=-self.n_links / 10, high=self.n_links / 10, size=2)
+    #         if np.linalg.norm(self.goal) < self.n_links / 10:
+    #             break
+    #     qpos[-2:] = self.goal
+    #     qvel = self.init_qvel + self.np_random.uniform(low=-.005, high=.005, size=self.model.nv)
+    #     qvel[-2:] = 0
+    #     self.set_state(qpos, qvel)
+    #     self._steps = 0
+    #
+    #     return self._get_obs()
+
     def _get_obs(self):
         theta = self.sim.data.qpos.flat[:self.n_links]
+        target = self.get_body_com("target")
         return np.concatenate([
             np.cos(theta),
             np.sin(theta),
-            self.sim.data.qpos.flat[self.n_links:],  # this is goal position
-            self.sim.data.qvel.flat[:self.n_links],  # this is angular velocity
-            self.get_body_com("fingertip") - self.get_body_com("target"),
-            # self.get_body_com("target"),  # only return target to make problem harder
+            target[:2],  # x-y of goal position
+            self.sim.data.qvel.flat[:self.n_links],  # angular velocity
+            self.get_body_com("fingertip") - target,  # goal distance
             [self._steps],
         ])
 
-class ALRReacherOptCtrlEnv(ALRReacherEnv):
-    def __init__(self, steps_before_reward=200, n_links=5, balance=False):
-        self.goal = np.array([0.1, 0.1])
-        super(ALRReacherOptCtrlEnv, self).__init__(steps_before_reward, n_links, balance)
-
-    def _get_obs(self):
-        theta = self.sim.data.qpos.flat[:self.n_links]
-        tip_pos = self.get_body_com("fingertip")
-        return np.concatenate([
-            tip_pos[:2],
-            theta,
-            self.sim.data.qvel.flat[:self.n_links],  # this is angular velocity
-        ])
-
-    def reset_model(self):
-        qpos = self.init_qpos
-        qpos[-2:] = self.goal
-        qvel = self.init_qvel
-        qvel[-2:] = 0
-        self.set_state(qpos, qvel)
-        self._steps = 0
-
-        return self._get_obs()
 
 if __name__ == '__main__':
     nl = 5
@@ -130,4 +140,4 @@ if __name__ == '__main__':
         if d:
             env.reset()
 
-    env.close()
\ No newline at end of file
+    env.close()