An arXiv paper describes a scene memory for robots and autonomous vehicles that runs on an FPGA rather than a desktop GPU. The authors argue that resource-constrained systems still need to localize a camera and render nearby views, and that this should happen on the device itself.

The work identifies scene-coordinate regression (SCR) and 3D Gaussian splatting (3DGS) as fitting approaches for such an edge scene memory. By building the memory around an FPGA, the design aims to bring these capabilities to platforms that cannot rely on a full GPU.

Since only the abstract was available, details about the implementation and evaluation are limited. The claim is that SCR and 3DGS are useful building blocks for this type of on-device scene memory, but the paper's specific contributions beyond that remain in the full text.