Baby steps towards physically based animation with reinforcement learning...

I've gone down this path before. Ended up writing reward functions to reward foot height over time within a periodic gait cycle. Use this just to get them stepping, then use a curriculum to turn this reward off and let the policy find a more natural gait pattern by penalizing energy. Had to penalize foot sliding too because they get lazy haha

Interesting, what happened if you didn't do all that? Just weird motion, or didn't learn at all?

Invest in your child's future success today with our Name Tracing Practice Board. Kid approved!

Super fun! I tried this in 2D once, though I used evolutionary methods - never managed to get RL working :/

Very cool! Can you explain a bit what you mean by evolutionary model? Randomizing weights in batches and move towards the winner? I'm using standard RL techniques (this one uses PPO)

I failed trying to make objects walk mechanically. Can you hear them screaming "kill me"?

Duuude! Finally someone doing (showing) game stuff related to AI models (that aren't just text/llm based). Massive potential here imo. The physics modelling could be huge too, but I guess the Teardown engine's physics is so optimised anyway it may not be needed! Exciting times!

Karl sims lives( I know he’s thankfully very much alive ;)

nice!

Delightfully creepy!