Abstract:

Traditional approaches for robotic arm motion planning assume static environments and often depend on prior knowledge about the shape and location of obstacles. Current scene mapping methods enable detailed scene reconstruction, but they are mostly suited for static scenes and they struggle to balance computational efficiency and fidelity. In this paper, we propose a unified framework for Gaussian-based scene rep- resentation and collision-free reactive robot control in unknown environments. We propose a real-time method for dynamic scene reconstruction from RGB-D images, enhancing the 3D Gaussian Splatting with key improvements under a fixed budget of Gaussians. Additionally, we introduce a technique for computing the signed distance function of the reconstructed environment using isotropic Gaussians, providing reduced computational complexity and smooth interpolation for computations of collision probability and reactive control. Our method demonstrates promising results in robot experiments across a range of environments, both in simulations and real-world settings.