Reproduce “Visual Navigation Transformer with Pose Attention” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Beiming Li, Jaime Romero, Jonathan Diller, Vijay Kumar, Alejandro Ribeiro

VNT-PA lets a robot navigate point goals in previously explored simulated indoor scenes by querying depth-image memories indexed by camera pose, not capture time. Rotary position embeddings make attention depend on pose differences, while the robot’s pose and goal query stored context. On HM3D validation, it achieves 93.3% success rate and 90.4% SPL; evidence is limited to simulation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.