[Paper Review] Semi-parametric Topological Memory for Navigation

Nikolay Savinov, Alexey Dosovitskiy+1 more|arXiv (Cornell University)|Mar 1, 2018

Advanced Image and Video Retrieval TechniquesComputer Science34 references55 citations

TL;DR

Introduces semi-parametric topological memory (SPTM), a graph-based memory with a retrieval network for navigation in unseen environments, achieving ~3x better success than baselines in maze tasks.

ABSTRACT

We introduce a new memory architecture for navigation in previously unseen environments, inspired by landmark-based navigation in animals. The proposed semi-parametric topological memory (SPTM) consists of a (non-parametric) graph with nodes corresponding to locations in the environment and a (parametric) deep network capable of retrieving nodes from the graph based on observations. The graph stores no metric information, only connectivity of locations corresponding to the nodes. We use SPTM as a planning module in a navigation system. Given only 5 minutes of footage of a previously unseen maze, an SPTM-based navigation agent can build a topological map of the environment and use it to confidently navigate towards goals. The average success rate of the SPTM agent in goal-directed navigation across test environments is higher than the best-performing baseline by a factor of three. A video of the agent is available at https://youtu.be/vRF7f4lhswo

Motivation & Objective

• Motivate navigation in unseen environments using a memory structure inspired by landmark navigation in animals.
• Propose a memory system that stores only connectivity (topology) and uses a neural retriever to localize within that memory.
• Train the retrieval and locomotion components in self-supervised fashion without manual labels or rewards.
• Demonstrate that SPTM enables efficient goal-directed navigation after ~5 minutes of exploration in novel mazes.

Proposed method

• Propose a memory architecture with a non-parametric graph G where nodes are locations and edges encode adjacency or visual similarity-based shortcuts.
• Introduce a parametric retrieval network R that computes similarity between observations to localize within G.
• Train R using self-supervised temporal proximity: pairs of observations are labeled as close if temporally near and far if separated by a margin.
• Build a locomotion network L that maps current and waypoint observations to action distributions, trained self-supervised from random exploration.
• During navigation, localize current and goal observations in G, compute a shortest path with Dijkstra, and select a reachable waypoint along that path using R.

Experimental results

Research questions

• RQ1 Can a semi-parametric topological memory enable reliable goal-directed navigation in unseen environments after limited exploration?
• RQ2 Does a topological memory with learned retrieval outperform metric-map or reactive baselines in generalization to new mazes?
• RQ3 How does the quality of localization and shortcut formation affect navigation performance?
• RQ4 Is self-supervised training of R and L sufficient for effective short-range navigation and planning?

Key findings

• SPTM achieves higher average navigation success than the best baseline by a factor of three on unseen mazes.
• With only 5 minutes of walk-through footage, the agent builds a topological map and navigates toward goals effectively.
• Ablation shows removing visual shortcuts dramatically degrades performance (Val-1:85%, Val-2:55%, Val-3:50% vs full:100%,100%,100%).
• Using per-frame localization or full memory texturing improves robustness, with full SPTM reaching 100% on Val-1 and Val-3 in validation.
• The shortest-path shortcuts reduce average path length dramatically (e.g., Val-3 path length drops from 990 to 155 steps).