Pranjal Mishra

Pranjal Mishra

Robotics Software Engineer · 3D Perception & Autonomy

ArtiTwinSplat: Interactable Digital Twin Reconstruction from RGB-D Videos

Master’s Thesis, CVG Lab ETH Zürich · arXiv

Tech stack

3D Gaussian Splatting · Nerfstudio · PyTorch · RGB-D reconstruction · Camera localization

Interactable digital twin reconstructed from a single RGB-D video.

Problem

Digital twins built with 3D Gaussian Splatting look photorealistic but are static: doors, drawers and other articulated parts can’t be moved. Robots need digital twins where they can interact with, and which can be built from ordinary RGB-D captures.

Overview

ArtiTwinSplat pipeline overview
Pipeline overview: RGB-D video → static and dynamic Gaussians → articulation estimation → interactable twin.
  • A photorealistic, interactable 3D scene reconstruction pipeline based on 3D Gaussian Splatting (3DGS), extended to dynamic articulated scenes via disentangled background/object Gaussians.
  • 4D RANSAC-based mobility fitting and an unsupervised RGB-D reconstruction framework to extract joint articulation priors for estimating object motion.
  • An extension of Nerfstudio’s WebRTC-based viewer that renders dynamic, articulated Gaussian Splat scenes live, so the digital twin can be viewed remotely in real time.

Results

Top: input RGB-D capture. Bottom: reconstructed twin with the articulated part moved.

My role

Master’s thesis project and first author of the workshop paper.

Presented as a lightning talk and poster at the ICRA 2026 Workshop on AI-Driven Automation and Robotic System Integration with Digital Twins, Vienna.