Sometimes DeepMimic isn't enough to track imperfect motion data robustly, e.g., clipping and ghost contacts between subject and objects.
In sirui-xu.github.io/InterMimic/ #CVPR2025, we present a method for learning from imperfect motion data, combining locomotion and object interaction.
This particular example is done using DeepMimic arxiv.org/abs/1804.02717, with very small changes to the reward and policy input. We have a related paper on motion cleanup coming out soon and looking forward to sharing it!
I think pdf links only provide enough information to link to show the "terrible" case. Guessing the "good" one shows up only when the pdf has been preprocessed. Not a Zotero user, but I worked on a bookmarklet providing similar functionality github.com/belinghy/PDF....
I really like this example. Besides the improved collision, the solution involved multiple steps:
1. L knee collides with R reducing lower body momentum
2. L falls early allowing R arm to swing over
3. L foot hooks R shin causing R to fall
4. R carries enough momentum to slide the right distance
Multiplayer motion cleanup is like solving a physics puzzle—agents learn to apply forces collaboratively to reproduce the desired motion. Recapturing with mocap is impractical—or even impossible depending on how closely the motion must match—but simulation makes it attainable.
Manually fixing broken multiplayer interaction animations is challenging. One of the main benefits of physics-based motion cleanup is that it is straightforward to extend to multiplayer—env code requires minimal modification, training code is unchanged, but training does take longer.
In character animation, animators often want to retain as much of the original motion as possible while fixing the motion artifacts, like sliding and clipping. Physics-based motion imitation systems like DeepMimic is very good at automatically detecting and removing these artifacts.
Quantifying motion quality is challenging, often subjective. While physics simulations don’t solve all problems, they offer a way to objectively improve quality by reducing physically implausible motions, for example:
