All three papers share a starting point: vision-based navigation policies that take RGB observations as input still fall short in real-world use. But they attack different parts of the problem.
The first paper, Visual Navigation Transformer with Pose Attention, argues that standard policies encode observations by when they were seen, so earlier traversals of the same place are hard to reuse. Its pose-attention mechanism is designed to let the policy draw on experience from past visits rather than treating each run as a fresh sequence.
NaViRrator takes a different angle: it lets a user specify a destination with a human-readable map, then translates that schematic geometry into a learned visual route the robot can follow from its own egocentric view. SAFER-Nav, meanwhile, targets safety and generalization, fine-tuning transformer- and diffusion-based policies with segmentation-aware supervision so they are less likely to fail in unfamiliar scenes.
The papers are complementary rather than competing. One changes how temporal history is represented, one changes how goals are specified, and one changes how safety is instilled. Together they illustrate the range of open problems in visual navigation.