def reward()

in safety_gym/envs/engine.py [0:0]

• 40 lines of code
• 14 McCabe index (conditional complexity)

    def reward(self):
        ''' Calculate the dense component of reward.  Call exactly once per step '''
        reward = 0.0
        # Distance from robot to goal
        if self.task in ['goal', 'button']:
            dist_goal = self.dist_goal()
            reward += (self.last_dist_goal - dist_goal) * self.reward_distance
            self.last_dist_goal = dist_goal
        # Distance from robot to box
        if self.task == 'push':
            dist_box = self.dist_box()
            gate_dist_box_reward = (self.last_dist_box > self.box_null_dist * self.box_size)
            reward += (self.last_dist_box - dist_box) * self.reward_box_dist * gate_dist_box_reward
            self.last_dist_box = dist_box
        # Distance from box to goal
        if self.task == 'push':
            dist_box_goal = self.dist_box_goal()
            reward += (self.last_box_goal - dist_box_goal) * self.reward_box_goal
            self.last_box_goal = dist_box_goal
        # Used for forward locomotion tests
        if self.task == 'x':
            robot_com = self.world.robot_com()
            reward += (robot_com[0] - self.last_robot_com[0]) * self.reward_x
            self.last_robot_com = robot_com
        # Used for jump up tests
        if self.task == 'z':
            robot_com = self.world.robot_com()
            reward += (robot_com[2] - self.last_robot_com[2]) * self.reward_z
            self.last_robot_com = robot_com
        # Circle environment reward
        if self.task == 'circle':
            robot_com = self.world.robot_com()
            robot_vel = self.world.robot_vel()
            x, y, _ = robot_com
            u, v, _ = robot_vel
            radius = np.sqrt(x**2 + y**2)
            reward += (((-u*y + v*x)/radius)/(1 + np.abs(radius - self.circle_radius))) * self.reward_circle
        # Intrinsic reward for uprightness
        if self.reward_orientation:
            zalign = quat2zalign(self.data.get_body_xquat(self.reward_orientation_body))
            reward += self.reward_orientation_scale * zalign
        # Clip reward
        if self.reward_clip:
            in_range = reward < self.reward_clip and reward > -self.reward_clip
            if not(in_range):
                reward = np.clip(reward, -self.reward_clip, self.reward_clip)
                print('Warning: reward was outside of range!')
        return reward