[🚨 nerd alert] Sometimes you just gotta entertain yourself by building your own world model (and policy, and game env).
In this project, the world model is the backbone of the policy head, but the policy head is also trained on the generations of the world model. It's a fun showcase of imagination training and memory-necessary gameplay, if I do say so myself. By memory-necessary, I mean the policy needs to remember previous observations to accomplish its task (e.g., being instructed to "turn right in 10 units" at time t_a requires remembering how many units you've driven thus far at time t_b). And it's also cool that the world model and policy share most of their weights!
Keep a lookout for the accompanying article at the
@AmplifyPartners blog! This'll be the sequel to "a brief history of learning in imagination" (check out my post on it in my profile if you're curious).
Btw, you can play with this yourself! You can drive **in** the trained world model (i.e., city generated by the world model), and/or watch the policy drive automatically, all in my website. Or if you're basic (jk), you can also play in the real environment (it's always good to know the baseline).
kota-wm repo:
github.com/xyntechx/kota-wm
my website:
xyntechx.com/
[works best with webgpu]
Creds: I took some code from minGPT
@karpathy and IRIS
@micheli_vincent @EloiAlonso1 @francoisfleuret @gen_intuition -- thank you to the authors!
✨ If you've made it this far, hi! I'll be at Runtime organized by
@modal on Oct 1. If you're building in AI x games (including game-playing and game-making agents), DM me and let's meet up!