Experiment: Key Seeker

Key Seeker is a follow-up to Green Dot. The task is is still fairly trivial: capture the yellow keys, then the green dot, as fast as possible. Avoid the red dot. Avoid the green dot until all of the yellow keys are captured.

This agent is not trained to exactly 100% success. If you watch long enough, you'll see an episode where the policy becomes confused and locks into a two state cycle until it times out.

Demo below.