SLAM uses explicit, geometry-based data structures — sparse feature-point maps, dense point clouds, 2D occupancy grids, 3D octrees, or voxel maps — that are directly readable by navigation and planning stacks. NeRF uses an implicit continuous neural field stored in MLP weights or hash-encoded feature grids. The NeRF map is compact, differentiable, and supports continuous surface modeling, but is not natively interpretable for classical path-planning.

No. NeRF requires accurate camera poses as input — it cannot independently localize the robot. SLAM's primary function is real-time pose estimation through feature tracking or scan matching, making pose estimation the cornerstone of the system. This is the primary reason early NeRF-SLAM integration systems such as iMAP and NICE-SLAM relied heavily on the NeRF rendering loss as a pose optimization signal rather than traditional feature matching.

Training a NeRF model from scratch is computationally expensive, typically requiring offline optimization over many seconds to minutes, far exceeding real-time requirements. The computational load of ray-marching-based rendering at map-building speed is excessive for embedded robotic platforms. However, advances such as multi-resolution hash encoding (Instant-NGP), triplane hash encoding, and octree-guided ray sampling have significantly reduced this gap and now enable near-real-time NeRF mapping.

Dynamic objects corrupt SLAM pose estimation via feature mismatches and contaminate NeRF's static scene assumption. The leading solution is segmenting and masking dynamic objects using SLAM's semantic pipeline before NeRF training. Systems such as Orbeez-SLAM use ORB-SLAM2 to provide initial poses, enabling fast NeRF-SLAM without pre-training and achieving real-time inference.

Novel view synthesis is NeRF's ability to produce photorealistic renderings at arbitrary viewpoints and arbitrary resolution — including perspectives the robot never physically visited. ULSAN Institute of Science and Technology demonstrated a system that generates "flip images" simulating unseen perspectives of SLAM objects and refines them via bundle adjustment, then applies a statistical NeRF model accounting for uncertainty — enabling scene coverage in areas the robot never physically visited.

Researchers are abandoning pure NeRF-only or pure SLAM-only systems in favor of hybrid architectures in which SLAM provides real-time pose estimation while NeRF handles dense scene reconstruction. Emerging trends include 3D Gaussian Splatting as an alternative to NeRF for dense SLAM, multi-agent NeRF SLAM with enhanced loop closure, and VLA (vision-language-action) model integration into SLAM for embodied intelligence.