Tag: the episode 6 robot (the chaser)
The green robot from episode 6 of A Million Tries ("AI learns to catch"): it learned to chase a hand-written blob, then got a memory of which way the blob went when it lost sight of it. About 853,000 practice rounds.
Catches my smartest hand-written blob (hides behind obstacles, keeps off the walls) 80% of the time, the simple "run away" script 96%.
Then the blob got a brain (episode 7): the first AI blob escaped this robot 91% of the time.
Sees (27 numbers): 16 distance lasers in a full circle (up to 30 m), its own speed and position, whether it can see the other player, where the other one is (or was last seen) and how it's moving (or was, when last seen), how long since it last saw it, and the time left. Obstacles block its view: no seeing through walls.
Does: an acceleration (2 numbers), 10 times a second. The robot is a little faster; the blob turns twice as hard.
Brain: 27 -> 256 -> 256 (tanh) -> 2 (+ a value head used only in training). Gaussian policy; the best-guess move is the mean (that's how it plays in the videos and the browser game).
Training: PPO, written from scratch in JAX, on a laptop CPU. The game: random 40 x 40 m arenas with pillars and walls, 20-second rounds, catch = touching. Rewards: +1 for a catch (the blob: for lasting the 20 s), -1 the other way round, a small cost per second, and a tiny hint per metre the gap shrinks (robot) or grows (blob).
Weights:
model.safetensors, PyTorch layout (body.0.weight, ...,policy.weight);config.jsonhas the training and game settings. Every number below comes from 192 test arenas no AI ever trained in, and that weren't used to choose this checkpoint.
Play against it (or any generation) in your browser: https://huggingface.co/spaces/amilliontries/beat-my-ai-tag
From the YouTube channel A Million Tries: https://www.youtube.com/@amilliontries
- Downloads last month
- 9