# Emergence of Maps in the Memories of Blind Navigation Agents

**Authors:** Erik Wijmans, Manolis Savva, Irfan Essa, Stefan Lee, Ari S. Morcos, Dhruv Batra  
**arXiv:** [2301.13261](https://arxiv.org/abs/2301.13261)  
**Date:** 2023-01-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.13261v1/

## Key points

- Organisms build and maintain internal spatial representations, or maps, of their environment
- Artificial intelligence (AI) navigation agents may also develop implicit maps
- AI researchers can design an agent's perceptual system and learning paradigm to eliminate alternative navigation mechanisms
- "Blind" agents with limited sensing capabilities can be trained to perform PointGoal navigation using reinforcement learning
- Blind agents are remarkably effective navigators in new environments with a success rate of approximately 95%
- Blind agents demonstrate the utilization of memory over long time horizons, remembering around 1,000 steps of past experience within a single episode
- Maps and collision detection neurons emerge within blind agents' internal representations of the surroundings as they navigate
- These emergent maps are selective and task-dependent
- The findings strengthen the understanding of mapping as a fundamental mechanism for navigation in both biological and artificial intelligent agents.

## Summary

In the field of animal navigation research, it is widely accepted that organisms build and maintain internal spatial representations, or maps, of their environment. This raises the question of whether machines, specifically artificial intelligence (AI) navigation agents, also develop implicit maps. Answering this question positively would not only explain the recent phenomenon observed in literature where seemingly map-free neural networks achieve impressive performance but also provide stronger evidence for mapping as a fundamental mechanism for navigation by intelligent embodied agents, regardless of whether they are biological or artificial. Unlike animal navigation, AI researchers have the ability to carefully design an agent's perceptual system and control the learning paradigm to eliminate alternative navigation mechanisms. In this study, the authors train "blind" agents with limited sensing capabilities restricted only to egomotion (self-motion) and no other forms of sensing. The objective is to teach these blind agents how to perform PointGoal navigation which involves reaching specific coordinates (Δx, Δy), using reinforcement learning. The agents used in this experiment consist of navigation-agnostic components such as fully-connected and recurrent neural networks. The experimental setup intentionally avoids any inductive bias towards mapping. Despite these challenging conditions, the results reveal several surprising findings about blind agents' navigational abilities. Firstly, blind agents prove to be remarkably effective navigators in new environments with a success rate of approximately 95%. Secondly, these agents demonstrate the utilization of memory over long time horizons by remembering around 1,000 steps of past experience within a single episode. This memory capacity enables them to exhibit intelligent behaviors such as following walls, detecting collisions and taking shortcuts. Furthermore, as blind agents navigate their environment maps and collision detection neurons emerge within their internal representations of the surroundings. These emergent maps are selective and task-dependent; for instance the agent tends to "forget" exploratory detours that are not relevant to its current goal. Overall this paper does not introduce new techniques to the AI audience however it presents a surprising finding that blind agents with limited sensing capabilities and no explicit bias towards mapping are able to navigate effectively and develop internal maps. This insight strengthens the understanding of mapping as a fundamental mechanism for navigation in both biological and artificial intelligent agents providing valuable implications for the design and development of AI navigation systems.