Reproduce “Data Scaling for Navigation in Unknown Environments” 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

Lauri Suomela, Naoki Takahata, Sasanka Kuruppu Arachchige, Harry Edelman, Joni-Kristian Kämäräinen

The authors study whether crowd-sourced demonstrations from more locations help sidewalk robots navigate unfamiliar places without maps. They train a policy to use camera images and noisy goal directions, and test it in four environments. With total training data fixed at 128 hours, doubling locations corresponds to about 15% fewer segment failures, as estimated by a fitted relationship.

Key insights

  • Geographical diversity in training data is a more critical factor for achieving robust generalization in point-goal navigation than simply increasing the total quantity of data from limited environments.
  • A power-law relationship exists, where doubling the number of distinct training locations consistently reduces navigation failures by approximately 15%, indicating continuous benefits from broader diversity.
  • Simpler, regression-based policy architectures are more robust and effective than complex generative or sequence models when trained on noisy crowd-sourced data and deployed remotely.

Summary by alphaXiv.

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