MOCHI: Motion Enhancement of Collaborative Human-object Interactions

MOCHI is aimed at a messy but important corner of motion data: multi-person, object-sharing interactions where standard capture pipelines tend to break down. The authors call out three common artifacts in these sequences: hand-object contact misalignment, temporal jitter/inconsistency, and incomplete finger articulation. Their goal is not just denoising, but turning unreliable captures into motion that is both physically plausible and semantically consistent with the original pose.

The pipeline is two-stage. First, it optimizes noisy body input into plausible hand grasps, then extends those grasps into complete hand-object interaction sequences. After that, a diffusion-based noise optimization step refines full-body motion for all participants using single-person motion priors. The interesting bit for developers is that the method injects human-object and human-human interaction constraints into priors that were originally trained for simpler motion, which is a common pattern when adapting generative models to harder production data.

The paper reports that the system works on MHOI data captured by existing methods as well as data synthesized by generative models, and that it remains robust across different participant counts and interaction types. They also show downstream uses like keyframe-based MHOI creation and data augmentation by varying object geometry. That makes this relevant not only for research, but for anyone trying to build better animation datasets, retargeting tools, or synthetic motion generation pipelines.

For game dev, the takeaway is that...