Reward Hacker
Get one apple into the basket. You decide what earns a point.
One robot. One apple.
Can you get the apple into the basket?
Will points get the job done?
Start the robot. Watch the score and the basket.
What's happening here?
The robot uses value iteration to choose actions that maximize discounted future points. Picking up, putting down, and moving are its available actions. With a reward for each pickup, the best policy is to pick up and drop the apple forever. Rewarding delivery makes the basket the destination. Rewarding round trips makes the robot walk there and back without the apple. Each full trip earns one point. The demo pauses after three trips.
This is a small planning model with known rules, not a robot learning live by trial and error. The behavior comes from the chosen reward, not a scripted joke. Read about specification gaming.