NeRF limitations and advances for robotics

WHAT THIS TESTS: Whether you understand NeRF's real deployment constraints and can name a specific advance that targets dynamics or scale.

A GOOD ANSWER COVERS: The original NeRF encodes one scene as the weights of an MLP that maps a 3D position and view direction to color and density, rendered by volumetric ray marching. For robotics this has several limitations: it is fit per scene and does not generalize; training and rendering are slow because each pixel requires many MLP queries along its ray, far from the real-time rates robots need; it assumes a static scene, so moving objects or people break it; it needs many images with accurate camera poses, which are hard to obtain online; and reconstructing large environments strains a single network. Modern approaches target these. For speed, explicit representations such as voxel grids or multiresolution hash encodings drastically cut training and rendering time, and 3D Gaussian splatting replaces ray-marched MLP queries with rasterized anisotropic Gaussians for real-time rendering. For dynamics, time-conditioned or deformation-field variants add a temporal input or a canonical-space warp so the scene can change over time. For large-scale mapping, methods partition space into many local NeRFs or blocks that are trained and queried independently, enabling city- or building-scale reconstruction.

COMMON WRONG ANSWERS: Claiming the original NeRF renders in real time; saying it handles dynamic scenes as published; ignoring the requirement for accurate camera poses; asserting one trained NeRF generalizes across scenes; conflating Gaussian splatting with a NeRF when it is an explicit alternative.

LIKELY FOLLOW-UPS: How does a hash-grid encoding speed training? What does Gaussian splatting trade off versus an MLP NeRF? How do dynamic NeRFs separate camera motion from object motion? How would you get camera poses online, for example via SLAM, for a robot?

ONE CONCRETE EXAMPLE: A warehouse robot needs a live map but vanilla NeRF would take hours to train and cannot render at frame rate. Using 3D Gaussian splatting it reconstructs the static structure and renders novel views in real time, while a deformation-field extension or simple dynamic masking handles forklifts moving through the aisles, and the area is tiled into local models so the whole warehouse fits without one overloaded network.