SplatCtrl: 3D Gaussian Splatting enables reactive, collision-free robot control

Robotic manipulators excel in structured settings but struggle in unstructured, dynamic environments. SplatCtrl, introduced by Siddarth Jain and Ho Jin Choi, tackles this by combining real-time 3D scene reconstruction with reactive motion planning. The framework builds on 3D Gaussian Splatting (3D-GS), incorporating a hybrid voxel-based filtering and dynamic Gaussian relocation strategy to efficiently reconstruct scenes from RGB-D streams while adapting to environmental changes. For safe control, the authors derive continuous signed distance functions from isotropic Gaussians, yielding stable, differentiable collision probability estimates that bridge classical distance fields with modern implicit representations. These metrics are integrated into control barrier functions, forming a perception-action coupling framework that generates smooth, real-time motion in response to scene changes.

Experimental validation spans simulation, physical robot platforms, and shared human-robot workspaces, demonstrating SplatCtrl's effectiveness in uncertain, dynamic environments. The system achieves integrated scene reconstruction and reactive control without pre-mapping, handling obstacles and human movements in real time. By grounding reactive motion in a probabilistic, differentiable representation, SplatCtrl marks a step toward safer, more adaptable robotic manipulators for human-centric settings. The paper appears at the 2026 International Conference on Robotics and Automation (ICRA).