Led by Professor Kyeongbo Kong, the team identified that existing Dynamic Gaussian Splatting methods struggle to generalize across diverse scenarios. Their solution introduces two distinct frameworks: MoE-GS, which trains separate dynamic models and blends them through learned routing, and MoDE, which integrates deformation experts during joint optimization. By moving away from a one-size-fits-all model, these systems allow the AI to select the most effective representation for specific regions and time steps.

This research, published in the IEEE Transactions on Pattern Analysis and Machine Intelligence, demonstrates that combining specialized experts significantly improves reconstruction quality in complex environments. The team expects this adaptive approach to provide a foundational shift for robotics, autonomous systems, and digital twins, where the ability to interpret heterogeneous motion is critical for natural interaction with the physical world.