"EPG takes a step toward agents that are not blank slates but instead know what it means to make progress on a new task, by having experienced making progress on similar tasks in the past."
Can someone explain to me how they take a step? It seems like they just use random search define a loss function for the sub-policy to optimize against. Is it because the loss function is "learned" over the sequence of actions, making it adaptive?
Parametrize your loss function and wrap a normal policy optimization with a random search to find a better loss function. Don't call it "random search," call it "evolution strategies" to make it sound sophisticated.
Would someone here know how to go about recreating a physics sandbox using a virtual robot arm with cubes in a game engine editor like Unity/UE4 where we'd be able to apply ML?
You certainly could, and gamedev tools are better than ever before at modelling real world physics. However, I would submit that if you want to simulate a robotic arm, it would be better to use tools specifically designed for that purpose. There are lots of reasonable-fidelity simulators of real robot arms which work with Gazebo, and by using the larger ROS ecosystem, you can also process simulated camera or depth camera data using standard pipelines, which will also be truer to life than using Unity.
