Robotics · 2025

Vision-Guided Differentiable Physics for Robotic Manipulation

Research & Engineering

End-to-end perception-to-control framework that uses 3D Gaussian Splatting and differentiable physics to identify object properties and plan contact-rich robotic manipulation.

Highlights

  • Integrated 3D Gaussian Splatting with NVIDIA Warp to build a differentiable perception–physics loop for robotic manipulation.
  • Used Isaac Lab to generate synthetic RGB-D data with ground-truth dynamics for system identification and planning.
  • Recovered latent physical parameters such as friction and mass from visual gradients and stored them in a neural scene-graph for long-horizon planning.