Point cloud vs voxel grid vs NeRF

WHAT THIS TESTS The interviewer wants a trade-off analysis grounded in robotics constraints, not a recital of definitions. The right representation depends on whether you prioritize geometry queries, memory, or photorealism.

A GOOD ANSWER COVERS Point clouds: a set of 3D points, often directly from depth sensors or LiDAR, compact for sparse scenes and fast to capture, but unstructured with no explicit connectivity, so collision checking needs spatial indexing like a KD-tree and surfaces are not watertight. Voxel grids: space discretized into cells storing occupancy or signed distance, regular and trivially queryable for collision and path planning, but memory grows cubically with resolution, mitigated by octrees or hashing. NeRF: an implicit neural function mapping position and direction to color and density, giving high-quality novel views and continuous geometry, but rendering requires many network queries so it is slow, and extracting explicit geometry for fast collision checks is indirect.

COMMON WRONG ANSWERS Saying NeRF is always best because it looks photorealistic, ignoring inference cost and the difficulty of collision queries. Or claiming voxels are obsolete, ignoring their planning advantages. Or treating point clouds as full surface models.

LIKELY FOLLOW-UPS How do octrees reduce voxel memory. How would you do collision checking against a NeRF, perhaps via density thresholding or distilling to an occupancy field. When is a signed distance field preferable. How does sensor type bias the choice.

ONE CONCRETE EXAMPLE For a robot navigating a room, you maintain a voxel occupancy grid via OctoMap for fast collision-free planning, since checking whether a cell is occupied is a direct lookup. You fuse incoming depth point clouds from the sensor into that grid frame by frame, using an octree so memory stays manageable at fine resolution. Separately, you can train a NeRF offline for high-quality visualization or photorealistic simulation, accepting its slow rendering because it is not on the real-time control path. Each representation is chosen for the task it serves best rather than forcing one to do everything.