>10×
single-instance speedup
over PhysTwin on embedded and desktop GPUs
ECCV 2026
One physical twin. Many parallel rollouts. One GPU.
Boba: Batched Simulation for Physics-Based Gaussian Digital Twins
1 University of Illinois Urbana-Champaign
2 Columbia University
3 Meta Reality Labs Research
Boba co-designs physics, Gaussian deformation, and rendering for batched execution—turning one reconstructed deformable object into many independent, high-fidelity simulation rollouts.
16.2×
aggregate throughput
over optimized single-instance execution on RTX 4090
13.2×
lower XR latency
Boba-Distributed versus PhysTwin on Orin
26×–>2,000×
faster predictive control
from rope to matched self-collision cloth planning*
Overview
High-fidelity twins should scale like simulations.
Existing Gaussian physical twins deliver realistic appearance and dynamics, but execute one instance at a time. Boba makes parallel rollouts a first-class systems problem.
Abstract
Digital twins replicate the appearance and physical behavior of real-world objects for interactive simulation in XR, robotics, and gaming. While neural representations such as 3D Gaussians enable high-fidelity twins reconstructed from images, existing systems simulate only one instance at a time and barely reach real-time performance, making large-scale rollout evaluation impractical.
We present Boba, the first batched simulator for physics-based Gaussian digital twins. Boba separates the static twin template from dynamic simulation state and co-designs physics, deformation, and rendering/visualization for batched execution. Compact surrogate spring–mass models, mixed-precision Gaussian skinning, memory-efficient execution, and shared-memory-aware batching reduce compute, memory traffic, and synchronization overhead.
Boba achieves over 10× single-instance speedup on XR-class hardware, scales to 3,310 FPS on RTX 4090, and substantially accelerates predictive control for deformable-object manipulation—supporting both power-constrained XR and scalable robot planning.
Method
A model–system co-design from physics to pixels.
Boba optimizes the complete digital-twin path and adds batching-aware memory layouts and scheduling where conventional process replication breaks down.
Compact physics
Surrogate spring–mass models reproduce dense-system behavior with substantially less compute and memory.
Efficient deformation
Mixed-precision Gaussian skinning and memory-traffic-aware execution accelerate per-frame updates.
Template sharing
Static twin assets are stored once while each simulation retains an independent dynamic state.
Contention-free batching
Shared-memory-aware layouts and atomic-free scheduling turn GPU capacity into sustained throughput.
Deployment variants
Three configurations. One optimized core.
Boba-Local
Complete single-instance execution on an edge device or workstation. On XR-class Orin, it reaches 30.8 FPS at 32.5 ms—a 10.2× latency improvement over PhysTwin.
Boba-Distributed
Server simulation and skinning with edge rendering. On XR-class Orin, it achieves 39.8 FPS at 25.1 ms—a 1.3× latency improvement and 22.2% lower device dynamic power than Boba-Local.
Boba-Batched
A shared GPU pipeline over many independent states. On RTX 4090, it reaches 3,309.9 FPS aggregate throughput—16.2× optimized single-instance execution on the same GPU.
Results
Fast locally. Power-efficient at the edge. Scalable in batches.
Boba-Distributed gives the best XR edge latency–power trade-off, Boba-Local maximizes raw single-instance desktop speed, and Boba-Batched maximizes aggregate throughput.
3,309.9
FPS
Boba-Batched on RTX 4090
Average of the best sustained aggregate throughput over full runs across evaluated scenarios.
39.8
FPS · 25.1 ms
Boba-Distributed for XR
1.3× latency improvement over Boba-Local on XR-class Orin, reducing end-to-end latency from 32.5 ms to 25.1 ms.
−22.2%
Device dynamic power
Boba-Distributed on XR-class Orin
4.76 W versus 6.11 W for Boba-Local on the same edge device.
204.5
FPS
Boba-Local on RTX 4090
12.1× faster than the same-hardware PhysTwin baseline.
30.8
FPS
Boba-Local on XR-class Orin
Real-time local execution at an approximately 15 W embedded compute budget.
Fidelity
Acceleration without abandoning the twin.
Across the full recorded interaction, Boba remains visually consistent with real observations and the PhysTwin baseline.
Applications
Planning, interaction, and learning.
The main paper and supplement demonstrate Boba across robot planning, immersive XR interaction, and reinforcement learning.
Immersive XR · Supplement
A complete interactive 3D application.
A Quest 3 user manipulates a simulated Gaussian physical twin inside a stereo reconstructed scene. The workstation handles simulation, skinning, rendering, composition, and application logic while streaming the composed views to the headset.
Download demo videoCloth planning
Target and execution
Rope planning
Target and execution
Model-based motion planning · Main + supplement
26×
faster rope predictive control
The supplement shows cloth and rope plans from both robot and third-person views. Parallel rollouts reduce the wall time of PhysTwin-based model predictive control; the paper also reports more than 16× for cloth without self-collision and an estimated greater-than-2,000× speedup with self-collision enabled.
Read the application study
MPC accuracy · Supplement
More rollouts without losing planning accuracy.
Boba matches full-PhysTwin rollout target CD in substantially less planning time; under the same time budget, PhysTwin evaluates fewer rollouts and yields higher target error.
RL training · Supplement
Higher-throughput learning from the same task.
Under identical PPO settings, Boba produces more interactions per wall-clock time and reaches higher reward and success earlier on the rope-reaching task.
Meet Boba
The cat behind the name.
Boba is Yihan’s cat and has supervised him throughout this project.
Resources
Paper, implementation, and citation.
BibTeX
Cite Boba
@inproceedings{pang2026boba,
title = {Boba: Batched Simulation for Physics-Based Gaussian Digital Twins},
author = {Pang, Yihan and Jiang, Hanxiao and Kondguli, Sushant and
Adve, Sarita and Wang, Shenlong},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026}
}* The greater-than-2,000× cloth result is an estimate reported for matched self-collision; see the paper for protocol and qualification.