Average success rate for five real-world robotic experiments for inserting an object in the frame. We use a control frequency of 15\,Hz, and one time-step corresponds to a single control action. Thus our method learns the task of insertion in less than four minutes of real-world training ($3000/(15*60) \approx 3.4$ minutes). It can be seen that both the proposed extensions work better than the original Residual TD3 method. This can be attributed to the fact that adding recurrence to the TD3 formulation can help in faster learning. Additionally, using the decay factor adds implicit constraints on the algorithm that correcting trajectory is more important as the robot is closer to the goal.