Mosca do Lule

A simulated fruit fly (DeepMind and HHMI Janelia's flybody, in MuJoCo), wearing a tiny red cap, trained with reinforcement learning to vote on a Brazilian electronic ballot box. It stands on the table, presses 1, then 3, waits for the candidate to show up on screen and hits CONFIRMA with its front leg. PIRIRIRI.

Everything is in CGS units like flybody itself: the fly is about 3.7 mm long and weighs about 1 mg. The clean vote showed up around 3.8M PPO steps, after a curriculum on the starting position and a penalty for wrong keys. Yes, it's a meme. It's also the most I've learned about reward shaping.