An on-policy reinforcement learning algorithm (Proximal Policy Optimization) and an off-policy algorithm (Hindsight Experience Replay) were used to train a policy on the existing OpenAI Gym environment with the Fetch Robot. A custom environment was also created to model the specific robotic equipment used in the labs at Northeastern by swapping the Fetch Robot with a UR3e robotic arm. The same two reinforcement algorithms were again used to train a policy on the custom environment for comparison. A Gazebo/ROS environment was also created to emulate the OpenAI Gym simulated environment.
  As shown in the above presentation, meaningful learning was achieved with the Hindsight Experience Replay (HER) algorithm. It was also shown that an on-policy algorithm such as Proximal Policy Optimization (PPO) is not suited for this type of complex environment with sparse rewards. The largest conclusion that was realized from this work is how important the choice of algorithm and reward function are. It also became readily apparent how closely these two choices are to each other. Some rewards functions may work well with certain algorithms but might work poorly with other algorithms.