For a literature review of Artificial Stupidity: https://arxiv.org/abs/1808.03644

For more information about why meta-learning (and in particular meta-RL) is useful see this post: https://bair.berkeley.edu/blog/2017/07/18/learning-to-learn/

The first applications appear really simple because they illustrate cognitive abilities that imitate planning/model-based RL related to neuroscience/psychology.

It doesn't mean that meta-RL won't scale up with more computation (see http://www.incompleteideas.net/IncIdeas/BitterLesson.html).

Hum, have not tried that but I guess the robotic part and the computer vision would be pretty standard, so it would be feasible.

PS:They even have robots playing real Pong! https://www.youtube.com/watch?v=tIIJME8-au8

Both the Keras one and the one from spinning up can be called "vanilla policy gradients". The one in Keras is closer to REINFORCE, and the one from spinning up use actor-critic and a multilayer perceptron.

Thanks! I would totally recommend Spinningup, because of the guide and the docs.