Emergence of Maps in the Memories of Blind Navigation Agents
The DOI matches its Crossref record; 33 of the 33 references it cites were found in a registry or an index; 9 of the 42 sources its claims cite were checked for support.
Abstract
Decades of research into intelligent animal navigation posits that organisms build and maintain internal spatial representations (or maps) 1 of their environment, that enables the organism to determine and follow task-appropriate paths (Epstein, Patai, Julian, & Spiers, 2017; O'keefe & Nadel, 1978; Tollman, 1948). Hamsters, wolves, chimpanzees, and bats leverage prior exploration to determine and follow shortcuts they may never have taken before (Chapuis & Scardigli, 1993; Harten, Katz, Goldshtein, Handel, & Yovel, 2020; Menzel, 1973; Peters, 1976; Toledo et al., 2020). Even blind mole rats and animals rendered situationally-blind in dark environments demonstrate shortcut behaviors (Avni, Tzvaigrach, & Eilam, 2008; Kimchi, Etienne, & Terkel, 2004; Maaswinkel & Whishaw, 1999). Ants forage for food along meandering paths but take near-optimal return trips (Müller & Wehner, 1988), though there is some controversy about whether insects like ants and bees are capable of forming maps (Cheung et al., 2014; Cruse & Wehner, 2011).
Overview
CoChat Summary
TL;DR
Blind PointGoal navigation agents trained with only egomotion sensing achieve high success and build implicit, task-dependent map-like memory representations. Analyses of memory transplants, collision probes, and occupancy decoding show that mapping emerges from goal-driven navigation without explicit mapping supervision or architectural bias.
Key findings
- Blind agents reach goals in unseen environments with 95.1%±1.3% success and 62.9%±1.6% SPL, compared with a clairvoyant Bug algorithm at 100% success and 46.0 SPL and a sighted depth agent at 94.0 success and 83.0 SPL.
- Memory is used over long horizons: memoryless agents achieve nearly 0% success, and performance does not saturate until memory contains about 1,000 steps, corresponding to 89.1±0.66 meters traveled.
- Collision information is linearly decodable from memory at 98% accuracy; the top 10 of 3072 neurons still achieve 96% collision-vs-not accuracy, reducing dimensionality by 99.7%.
- Transplanting a trained agent’s final memory into a blind probe improves SecondNav(S→T) SPL to 85.0±0.16 and SecondNav(T→S) SPL to 84.8±0.22, versus controls near 71–72 SPL; occupancy maps decode at 32.5% vs 12.5% IoU for trained vs untrained memory.
Methods
- Agents were trained for PointGoalNav in AI Habitat using Gibson and Matterport3D scenes, with only an idealized GPS+Compass egomotion sensor, four actions, a 3-layer 512-d LSTM, and DD-PPO/PPO; evaluation used 18 Matterport3D test scenes and 1008 episodes.
- Analyses included truncating memory length, training linear collision probes, transplanting agent memory into probe agents for SecondNav(S→T) and SecondNav(T→S), decoding occupancy grids, and decoding past/future positions.
- Performance was measured by Success and Success weighted by inverse Path Length (SPL); excursion analyses manually labeled 216 randomly sampled evaluation episodes.
Limitations
- The paper states that experiments use ground-based agents in digitized real houses, limiting them to a 2D manifold with strong structural priors, so generalization to settings like drones in forests is unclear; it also notes no actuation noise, only short episode-length exploration before memory reset, and no complete mechanistic account of map learning.
- The paper states that sighted depth-equipped agents did not provide convincing evidence of decodable metric maps, with mixed results: UntrainedAgentMemory had higher IoU (42.9% vs 40.1%) while TrainedAgentMemory had higher class-balanced accuracy (61.8% vs 53.1%).
- Evident from the text: probe evaluations remove episodes where the agent fails (about 5%), occupancy decoding is constrained to locations within 2.5 meters of visited areas, and excursion forgetting is quantified using manual labels on a limited sample of 216 episodes.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 3 Not valid · 18 Not checked · 111 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“More recently, the machine learning commu- nity has produced a surprising phenomenon – neural-network models for navigation that curiously do not contain any explicit mapping modules but still achieve remarkably high performance (Savva et al., 2019; Wijmans et al., 2020; Kadian et al., 2020; Chattopadhyay et al., 2021; Khandelwal et al., 2022; Partsey et al., 2022; Reed et al., 2022).”
What the sources say
- Not checked — no full text of the sourcecites Savva et al., 2019 source
- Not matchedcites Wijmans et al., 2020
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Supportedcites Kadian et al., 2020 source
The Navigation Models section (p. 6) describes a learned navigation agent with a visual encoder and LSTM policy, without an explicit mapping module. In the simulation comparison (p. 7), that model outperforms an oracle given a top-down map, supporting the claim that such models can achieve high navigation performance. (LLM, conf=0.90)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In a similar manner, convergent evolution (Kozmik et al., 2008), where two unrelated intelligent systems independently arrive at similar mechanisms, suggests that the mechanism is a natural response of having to adapt to the environment and the task.”
What the sources say
- Partly supportedcites Kozmik et al., 2008 sourceExperimental — accuracy not yet measured
The Discussion describes independently recruited lens components in cubozoan and vertebrate eyes as an example of convergence in eye evolution. It does not study intelligent systems or establish that a shared mechanism is a natural response to adapting to an environment and task; it also says the evidence does not unambiguously distinguish independent evolution from common ancestry. (LLM, conf=0.86)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: unclear — It is unclear whether it could be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Surprisingly, we find that agents trained under this impoverished sensing regime are able to navigate with near-perfect efficacy – reaching the goal with 95.1%±1.3% success rate (Table 1), even in situa- tions where the agent must take hundreds of actions and traverse over 25m.”
- Also stated in: § 1 Blind 95.1±1.3 62.9±1.6 · plain
“1 Blind 95.1±1.3 62.9±1.6”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1 Blind 95.1±1.3 62.9±1.6 · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Intriguingly further still, we find that surprisingly detailed metric occupancy maps of the environment (indicating free-space) can be explicitly decoded from the agent's memory.”
- Also stated in: § 4 MEMORY ENABLES SHORTCUTS · plain
“Learning navigation improves map prediction from memory.”
- Also stated in: § 4 MEMORY ENABLES SHORTCUTS · plain
“TrainedAgentMemory has a higher mean than UntrainedAgentMemory with p-value ≤ 10−5 (via Wilcoxon signed-rank test (Wilcoxon, 1992)).”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 4 MEMORY ENABLES SHORTCUTS · plain
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Indeed, we find that TrainedAgentMemory probe performs the same (within margin of error) on both SecondNav(S→T) and SecondNav(T→S) (Fig. 3B right column) – indicating that the memory is equally useful in both directions.”
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
72 more claims
Create a free account to see all 140 claims and to verify the 4 references still open.
- Probably not a substantial claim. 42
- It may state two separate claims as one. 29
- Not yet classified. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“Wijmans et al . [62] showed blind agents modeled using RNNs are capable of building map-like representations for the PointNav [45] task.”
From GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation (2024) · cites this paper as [62]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ REFERENCES
“SPLi = Successi · di min{di, li} (3)”
§ 1 Blind 95.1±1.3 62.9±1.6
“1 Blind 95.1±1.3 62.9±1.6”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 6 MAPPING IS TASK-DEPENDENT: AGENT FORGETS EXCURSIONS
“ŝt−k = fk(ht, ct)+st, k ∈ [1, 256].”
§ 6 MAPPING IS TASK-DEPENDENT: AGENT FORGETS EXCURSIONS
“Given ground truth location st+k, we evaluate the decoder via relative L2 error ||ŝt+k−st+k||/||st+k−st||”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4 MEMORY ENABLES SHORTCUTS
“85.0%±1.6% (probe)”
§ 3 TrainedAgentMemory 96.2±0.23 85.0±0.16 96.0±0.16 84.8±0.22
“3 TrainedAgentMemory 96.2±0.23 85.0±0.16 96.0±0.16 84.8±0.22”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ REFERENCES
“increase spatial resolution to [16, 96, 96].”
§ REFERENCES
“Specifically, after an initial CoordConv with an output channel- depth of 128, we use a series of 4 CoordUpConv-CoordConv layers where each CoordUpConv doubles the spatial dimensions (quadruples spatial resolution) and each CoordConv reduces channel-depth by half.”
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 11×
“Further still, our agents are composed of navigation-agnostic, generic, and ubiquitous architectural components (fully-connected layers and LSTM-based recurrent neural networks), and our experimental setup provides no inductive bias towards mapping – no map-like or spatial structural components in the agent, no mappin…”
BackgroundMethodInfluentialCognitive maps in rats and men
Tolman, E. C. (1948) · Psychological Review · doi:10.1037/h0061626 · Cited by 6,641 · cited 2×
“Decades of research into intelligent animal navigation posits that organisms build and maintain internal spatial representations (or maps) 1 of their environment, that enables the organism to determine and follow task-appropriate paths (Tolman, 1948; O’keefe & Nadel, 1978; Epstein et al., 2017).”
BackgroundThe Dance Language and Orientation of Bees
Frisch, K. V. (1993) · Harvard University Press eBooks · doi:10.4159/harvard.9780674418776 · Cited by 2,563
The Hippocampus as a Cognitive Map
RANCK, J. B. (1979) · American Journal of Psychiatry · doi:10.1176/ajp.136.10.1353 · Cited by 2,346
Habitat: A Platform for Embodied AI Research
Savva, M., Kadian, A., Maksymets, O., Zhao, Y., Wijmans, E., Jain, B., Straub, J., Liu, J., Koltun, V., Malik, J., et al. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00943 · Cited by 1,237 · cited 9×
“Specifically it seeks to learn a function Collided t = f (( h t , c t )) where ( h t , c t ) is the internal state at time t and Collided t is whether or not the previous action, a t 1 lead to a collision.”
BackgroundMethodInfluentialSelf-citationThe cognitive map in humans: spatial navigation and beyond
Epstein, R. A., Patai, E. Z., Julian, J. B., & Spiers, H. J. (2017) · Nature Neuroscience · doi:10.1038/nn.4656 · Cited by 921 · cited 1×
“Decades of research into intelligent animal navigation posits that organisms build and maintain internal spatial representations (or maps) 1 of their environment, that enables the organism to determine and follow task-appropriate paths (Tolman, 1948; O’keefe & Nadel, 1978; Epstein et al., 2017).”
BackgroundThe Dance Language and Orientation of Bees
Morse, R. A. (1994) · American Entomologist · doi:10.1093/ae/40.3.187a · Cited by 903 · cited 3×
“This perceptual system is deliberately impoverished to isolate the contribution of memory, and is inspired by blind mole rats, who perform localization via path integration and use the Earth’s magnetic field as a compass (Kimchi et al., 2004).”
BackgroundLumelsky, V. J., & Stepanov, A. A. (1987) · Algorithmica · doi:10.1007/bf01840369 · Cited by 699 · cited 4×
“To put this SPL in context, we compare it with ‘Bug algorithms’ (Lumelsky & Stepanov, 1987), which are motion planning algorithms inspired by insect navigation, involving an agent equipped with only a localization sensor.”
BackgroundMethodInfluentialPath integration in desert ants, Cataglyphis fortis
Müller, M., & Wehner, R. (1988) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.85.14.5287 · Cited by 677
On evaluation of embodied navigation agents
Anderson, P., Chang, A., Chaplot, D. S., Dosovitskiy, A., Gupta, S., Koltun, V., Kosecka, J., Malik, J., Mottaghi, R., Savva, M., et al. (2018) · arXiv · doi:10.48550/arxiv.1807.06757 · Cited by 622
Self-citation
Research works printed but not linked: 6
- Institute , S. R. ( 1972 ). Shakey: An experiment in robot planning and learning . Institute, S. R. (1972). Shakey: An experiment in robot planning and learning.
- Lockley , R. M. ( 1967 ). Animal navigation . Pan Books . Lockley, R. M. (1967). Animal navigation. Pan Books.
- Peters , R. ( 1976 ). Cognitive maps in wolves and men. Environmental design research, 2, 247--253 . Peters, R. (1976). Cognitive maps in wolves and men. Environmental design research, 2, 247--253.
- 10.5555/93002.93291 · doi:10.5555/93002.93291
- 10.5555/1121596 · doi:10.5555/1121596
- Wijmans , E. , Kadian , A. , Morcos , A. , Lee , S. , Essa , I. , Parikh , D. , ... Batra , D. ( 2020 ). DD-PPO: Learning near-perfect point-goal navigators from 2.5 billion frames . In Proceedings of the international conference on learning representations (iclr). Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., ... Batra, D. (2020). DD-PPO: Learning near-perfect point-goal navigators from 2.5 billion frames. In Proceedings of the international conference on learning representations (iclr).
Cited by
Learning robust autonomous navigation and locomotion for wheeled-legged robots
Lee, J., Bjelonic, M., Reske, A., Wellhausen, L., Miki, T., & Hutter, M. (2024) · Science Robotics · doi:10.1126/scirobotics.adi9641 · Cited by 252
The alignment problem from a deep learning perspective
Ngo, R., Chan, L., & Mindermann, S. (2022) · arXiv · doi:10.48550/arxiv.2209.00626 · Cited by 68
“Similarly, Banino et al. [2018] and Wijmans et al. [2023] identified representations which helped policies plan their routes when navigating, including in unfamiliar settings.”
BackgroundResilient Legged Local Navigation: Learning to Traverse with Compromised Perception End-to-End
Zhang, C., Jin, J., Frey, J., Rudin, N., Mattamala, M., Cadena, C., & Hutter, M. (2024) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra57147.2024.10611254 · Cited by 49
“RL-based local navigation has recently emerged as it is powerful for behavior learning [27]–[30], computationally efficient and suitable for mapless navigation [12], [13], [31]– [33].”
BackgroundMethodGOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation
Khanna, M., Ramrakhya, R., Chhablani, G., Yenamandra, S., Gervet, T., Chang, M., Kira, Z., Chaplot, D. S., Batra, D., & Mottaghi, R. (2024) · 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52733.2024.01549 · Cited by 23
“[62] showed blind agents modeled using RNNs are capable of building map-like representations for the PointNav [45] task.”
MethodSelf-citationNavigates Like Me: Understanding How People Evaluate Human-Like AI in Video Games
Milani, S., Juliani, A., Momennejad, I., Georgescu, R., Rzepecki, J., Shaw, A., Costello, G., Fang, F., Devlin, S., & Hofmann, K. (2023) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3544548.3581348 · Cited by 36
“More generally, it is considered fundamental to embodied biological intelligence [26, 59], making it of interest to cognitive scientists [56, 62] and researchers interested in intelligent behavior [87].”
Background
Metadata
- Authors
- Wijmans, E., Savva, M., Essa, I., Lee, S., Morcos, A. S., & Batra, D.
- Affiliations
- Georgia Institute of Technology (US)
- Simon Fraser University (CA)
- Oregon State University (US)
- FAIR, Meta AI
- Venue
- AI Matters
- Published
- 2023-06
- Volume
- 9
- Issue
- 2
- Pages
- 8-14
- ISSN
- 2372-3483 (ISSN-L)
- OpenAlex
- W4387476016
- Cited by
- 17
- References
- 27
- Citations by year
- 2026: 1 · 2025: 6 · 2024: 5 · 2023: 5
- FWCI
- 1.24
- Type
- journal article
- Language
- en
- Open access
- closed
- Locations
- AI Matters (publishedVersion)
- Topics
- Social Sciences › Psychology › Experimental and Cognitive Psychology › Spatial Cognition and Navigation
- Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Keywords
- spatial maps, animal navigation, blind navigation, memory, path planning, artificial agents
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Emergence of Maps in the Memories of Blind Navigation Agents. https://app.cochat.ai/papers/10.1145/3609468.3609471
- The paper
- Wijmans, E., Savva, M., Essa, I., Lee, S., Morcos, A. S., & Batra, D. (2023). Emergence of Maps in the Memories of Blind Navigation Agents. AI Matters. https://doi.org/10.1145/3609468.3609471
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.