Reproduce “Semi-parametric Topological Memory for Navigation” 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
Nikolay Savinov, Alexey Dosovitskiy, Vladlen Koltun
Researchers from ETH Zürich and Intel Labs developed the Semi-Parametric Topological Memory (SPTM) for autonomous navigation, which integrates a non-parametric topological graph with parametric deep networks to enable goal-directed movement using only visual observations. The system achieved a 100% success rate in novel 3D maze environments, tripling the performance of leading baseline methods.
Key insights
- Integrating a biologically inspired, non-metric topological memory graph with deep learning models enables robust navigation in complex, unseen 3D environments.
- Self-supervised training can effectively teach deep networks to perform visual similarity recognition and short-range locomotion, removing the need for explicit metric ground truth or labeled data.
- Explicitly constructing a non-parametric memory graph from observations, including visual shortcuts, significantly improves an agent's ability to generalize and navigate compared to unstructured deep memory (LSTMs).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.