Boba: Efficient and Scalable Physics-Based Simulation of Gaussian Digital Twins

DEMO

Yihan Pang, Hanxiao Jiang, Sushant Kondguli, Sarita Adve, Shenlong Wang

Boba is a systems framework for making Gaussian-based physical digital twins practical for XR and robotics. This demo showcases an interactive XR application built on Boba, highlighting how its co-designed system architecture enables responsive, physically grounded interaction with Gaussian digital twins.

More broadly, Boba supports both single-device and distributed deployment. In the single-instance XR setting, it delivers about a 10× end-to-end latency reduction across both edge-class and workstation-class platforms, reaches over 30 FPS on an XR-class edge-device proxy, and exceeds 150 FPS on a single RTX 4090. In a distributed edge–server configuration, Boba uses a workstation-class GPU server for simulation while the XR-class edge device handles visualization. Compared with running the full pipeline on the edge alone, this reduces end-to-end latency by 23% and device-side dynamic power by 22%.

Boba also supports large-scale batched simulation, scaling to over 6,000 simulation steps per second and up to 1,734 parallel instances on a single RTX 4090, and enabling MPC rollouts up to 2000× faster than the previous state of the art. Together, these capabilities and results position Boba as a strong systems foundation for physics-based Gaussian digital twins across XR and robotics.



Boba: Efficient and Scalable Physics-Based Simulation of Gaussian Digital Twins

Boba: Efficient and Scalable Physics-Based Simulation of Gaussian Digital Twins