Vision Robotic Manipulator Combined with Neural Radiance Field for Autonomous Real-time 3D Reconstruction

We present a new method to achieve accurate and real-time 3D scene reconstruction based on monocular cameras on a vision robotic manipulator platform. To realize this, we use the initialization part of SLAM and the latest advances in N eRF (Neural Radiance Field), Instant-NGP, and, most importantly, the information from the robotic manipulator to acquire real-time camera pose. This work uses keyframe images with camera real-time pose obtained from the visual robotic manipulator. Thus, the problem that instant-NGP needs images with pose are solved on the platform of the robotic manipulator, reducing the time of secondary processing of pure images and, consequently, realizing real-time 3D reconstruction.