Loading video...
Video Failed to Load
Whatever happened to training physics based character controllers using RL? This was all the rage a few years ago. Every clip here looks cool af.
29,075 views • 3 months ago •via X (Twitter)
11 Comments

still going strong. Check out our latest work GPC:

This is really cool 😍

Papers in this video:

it's still all the rage this year! Check out GPC, SMP, HIL, etc. These are experiments with GPC:

peopel who were doing this are now busy minting synthetic data en masse

We do this with robots with the @rek sim

If a virtual object knows how to behave in accurately modeled virtual space, it should know how to in reality Then apply a multiverse simulation to train it on all possibly outcomes Insane potential. Something im very interested in

RL slow, RL expensive, RL fails to generalize to edge cases etc

Character-control progress may depend more on real-world physics than prettier demonstrations.

Probably still eating too much resources. The people doing this research were mostly from AAA studios where resources are already tight. Also unlike traditional animation when this fails it fails in miserable ways. Indies could use this though.

Been a bit busy, but wanted to jump back here and mention I'm working on exactly this. There's no reason this shouldn't be the future of characters in games.

