Rapid Exploration for Open-World Navigation with Latent Goal Models
1. Paper overview
In one sentence: RECON combines compressed, context-conditioned goal sampling with topological memory to accelerate outdoor exploration and return navigation, while relying on heuristic reachability and online adaptation. e02e04e05e06e08
| At a glance | What to know |
|---|---|
| Research problem | Source description Given only a goal photograph, a robot must discover its location in an unfamiliar outdoor environment and subsequently return efficiently. Prior trajectories come from other environments; deployment supplies RGB observations without localization, a prebuilt map or known dynamics. The system is partially observed and commands continuous linear and angular velocities. e02e14 |
| Core mechanism | Author claim An information bottleneck compresses the goal relative to the current observation while retaining information useful for action and distance prediction. The authors argue that this representation suppresses irrelevant appearance variation and makes prior samples useful exploration targets. e04 |
| A key reported result | Visual goal discovery in unseen outdoor environments: RECON: 09:54 Exploration time (mm:ss), lower is better. Eight unseen real-world environments, one target RGB image per environment; methods use the same 20-hour offline dataset and online fine-tuning. ECR: 14:54; ViNG: 19:48; ANS: 21:00; PPO + RND: 21:18; InfoBot: 23:36. Table 1 favors RECON. These times imply a reader-calculated 33.6% reduction versus ECR, not a 50% time reduction. The caption says '50% faster.' Failure aggregation, trial counts and uncertainty are not specified; NTS is excluded after failing in every environment. e07e08 |
| Reading caution | Source description The method does not explicitly account for value of information. The authors leave guarantees for efficient stochastic-policy and bottleneck exploration to future work. e12 |
Core contributions
- Author claim
An information bottleneck compresses the goal relative to the current observation while retaining information useful for action and distance prediction. The authors argue that this representation suppresses irrelevant appearance variation and makes prior samples useful exploration targets. e04
Figure 2. Prior navigation experience supports exploration, and the resulting image memory supports return routes. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with panel (a), the user's goal photograph. In panel (b), the yellow point marks the start, the cyan trace shows exploration, white circles mark map nodes, and the red trace shows the discovered route. Panel (c) connects this route to the robot's own views. Panels (d) and (e) separate example training scenes from unseen test scenes. The overhead image is an illustration of the experiment; Section 3 specifies that the controller receives onboard RGB observations without ground-truth localization. Read the graphic as three stages: learn from earlier environments, explore a new one, then reuse its memory. e14e02e03e05e06e08
What it supports. The transferable component is local, goal-conditioned navigation knowledge. The topological map is built from experience in the new environment and makes that exploration reusable. This distinction explains why offline training and online environment discovery both belong to the method, rather than treating the result as navigation with a pre-existing map.
Where the evidence stops. These panels illustrate selected environments and trajectories. They do not establish a per-environment success distribution or show that the robot uses aerial imagery. Quantitative comparisons belong to Table 1.
2. Motivation
2.1 The problem and the proposed response
Given only a goal photograph, a robot must discover its location in an unfamiliar outdoor environment and subsequently return efficiently. Prior trajectories come from other environments; deployment supplies RGB observations without localization, a prebuilt map or known dynamics. The system is partially observed and commands continuous linear and angular velocities. e02e14
2.2 What this reading follows
A photograph tells a robot what its destination looks like, but not how to find it in a new environment. RECON addresses that gap by learning a compact relationship between the current view and a visual goal. Its decoder predicts actions and temporal distance; sampled latent goals turn that local navigation knowledge into exploratory behavior. A growing graph of encountered images then supports frontier selection and efficient return routes. The experiments test this system on a physical ground robot. Read the visuals to separate the learned representation, the memory it builds, the reported navigation gains and the limits of the supplied evidence. e02e04e05e06e08
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational action-policy and planning classification is supported. Figure 3 jointly predicts actions and temporal distances from a goal latent, while Algorithms 1–2 use an external topological memory. This is neither future-observation generation nor action extraction from a predicted world rollout; sharing a decoder does not establish the catalog's One Model world–action architecture. The recorded not-applicable architecture, prediction paradigm and quadrant are appropriate. e04e05e06
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Learn a goal representation that can drive motion
The key training decision is what the goal latent must preserve. A representation trained to reconstruct a photograph would have to retain appearance details even when they do not help the robot move. RECON instead makes the latent useful for predicting an action and the number of time steps to the goal. The current view conditions both encoder and decoder, so the latent describes the goal in context. Equation 1 expresses the prediction-versus-compression tradeoff; Equation 2 supplies the variational likelihood and KL objective. The standard normal prior also becomes an interface for generating exploratory subgoals. The authors argue that this produces feasible local goals, but the objective itself does not prove physical reachability for every prior sample. e03e04e05
Figure 3. The bottleneck retains goal information useful for actions and temporal distance. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the magenta arrows from goal observation O_g and current observation O_t into latent Z_t^g. Then follow the blue arrows: both the latent and current observation feed action A_t^g and distance D_t^g. Thus the goal is represented relative to the current view, while the decoder still receives current visual context directly. Equations 1–2 train this representation with an action/distance prediction objective and a compression penalty toward a normal prior. Sampling that prior supplies latent subgoals for exploration. The graphic specifies probabilistic dependencies; it does not supply layer counts, latent dimensionality or a reconstruction network. e04e05e13
What it supports. The same latent interface supports two behaviors: encode an observed goal to pursue it, or sample a latent to propose exploration. The learned decoder turns either latent into actions and a temporal-distance estimate. The model is trained for navigation relevance, without an image-reconstruction term in the displayed objective.
Where the evidence stops. The decoder box prints z_g^t, whereas its latent node and Equation 2 use z_t^g; the guide follows the latter notation. The arrows agree with the equations. Prior density remains a heuristic, not a reachability guarantee.
5.2 Turn a local controller into a persistent exploration system
Imagine that the supplied destination photograph is initially unfamiliar. Algorithm 1 evaluates the encoded goal under the prior. If its density exceeds epsilon, the robot pursues that goal directly. Otherwise, it uses the graph to find the least-explored nearby node. When already there, it samples a latent subgoal; when elsewhere, it first heads toward that frontier node. Each selected latent stays fixed for H steps, but actions are recomputed from incoming observations. The resulting trajectory expands memory and supplies fine-tuning data. Once the destination is discovered, Algorithm 2 changes the problem: associate start and goal with graph vertices, find a shortest path, and navigate through its image subgoals. Local prediction, physical execution, memory updates and route planning therefore have distinct roles. e05e06
5.3 Separate evidence for exploration, recall and robustness
Reader analysis: no single experiment establishes every part of the proposed explanation. Table 1 measures goal discovery and subsequent traversal, favoring the complete system under the reported protocol. Figure 6 removes graph construction and compares exploratory trajectories, which helps test the contribution of sampled-goal behavior. Table 2 then removes memory or changes goal generation on a subset of environments, revealing different costs for discovery and recall. Figure 5 adds evidence that previously learned navigation can survive selected changes in obstacles and appearance. Together these results support the system design, but do not uniquely identify compression as the cause of invariance. A stronger causal test would control online adaptation and training capacity while measuring representation behavior and physical success under matched perturbations. e07e08e09e10e11e04e05
5.4 Training and inference
During training
The navigation model uses supervised distance regression and goal-conditioned behavior cloning. Training uses 20 hours of offline data from environments distinct from testing. Self-supervised collection used time-correlated random walks over 18 months; time steps proxy distance, with relabeling deferred to missing Appendix A. e03e07
Equation 2 maximizes action-and-distance log likelihood with a beta-weighted KL penalty against a standard normal latent prior. During exploration Algorithm 1 updates encoder and decoder using accumulated new data for gamma epochs. Numerical hyperparameters and network/training details are deferred to absent Appendix B.2; no frozen-module scheme is specified. e04e05e13
During inference
Prior density is used as a confidence heuristic for goal feasibility. A high density suggests an in-distribution observation–goal relation; it is not an established probability of physical success. Local execution remains observation-conditioned despite holding the subgoal latent fixed. e05
Deployment includes adaptation during exploration. Later graph-based navigation reuses the explored environment through Algorithm 2. Neither stage synthesizes future images: the learned outputs are control actions and temporal distances, while the memory stores observations. e04e05e06
5.5 Implementation flow
- Encode a relative goal
The encoder p_phi conditions on both current observation o_t and goal image o_g. Its latent z_t^g feeds a decoder alongside o_t to predict action a_t^g and temporal distance d_t^g. Figure 3 contains direct current-observation links to both outputs; it is a probabilistic dependency diagram, not a detailed network specification. e04
- Choose where to explore
Algorithm 1 first tests whether the encoded goal has prior density above epsilon. If so, it pursues that goal. Otherwise, at the least-explored nearby node it samples a latent from the prior; away from that frontier it encodes the frontier image as its subgoal. Low node-selection counts define the frontier, with learned-distance thresholds restricting neighbors and arrival. e05
- Execute and remember
SubgoalNavigate holds the selected latent fixed for H steps while querying the decoder with successive observations. Actions drive the robot; predicted distances support graph edges. The rollout updates accumulated data, graph structure and visitation counts. Following exploration, current and goal images are associated with vertices and a shortest path supplies successive image subgoals. e05e06
6. Experiments & results
RECON learns how a visual goal relates to the robot's current view, predicts actions and temporal distance, and samples latent subgoals for exploration. A topological image memory supports frontier selection and later route reuse. Real-robot experiments favor this combination over random-action exploration, but missing appendices and limited statistical reporting constrain reproduction and causal interpretation.
6.1 Read the original evidence
Table 1. RECON leads the reported discovery-time, return-time and weighted-success comparisons. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the first timing column as time to discover the visually specified goal, and the second as time for a subsequent traversal using exploration experience. Both favor smaller values; SCT favors larger values. Section 5.1 describes eight unseen environments, one target image per environment, and a shared 20-hour offline dataset with online fine-tuning. The shaded ECR row is a close comparison: it retains the RECON-style graph but explores using random actions at the frontier. The other rows represent broader navigation or exploration alternatives. NTS has no row because the authors report that it failed to explore successfully in every environment. e07e08e05
What it supports. RECON reports 09:54 for discovery versus ECR's 14:54, and 00:26 for subsequent navigation versus 00:31. SCT rises from 0.73 to 0.92. From the reported discovery times, the reader-calculated time reduction is 33.6%; the caption's '50% faster' wording should not be read as a 50% reduction in elapsed time.
Where the evidence stops. The source does not give error bars, trial counts or an explicit rule for incorporating failed trials into timing summaries. SCT is weighted by completion time and must not be relabeled as raw success percentage.
Figure 5. A previously explored route remains usable in the illustrated obstacle and appearance changes. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Panel (a) supplies the goal image, and panel (b) shows the discovered path. The middle row introduces trash cans, traffic cones and a car; the bottom row changes lighting or weather, with labels Sunny, Rainy/Overcast and Twilight. Cyan curves show the presented routes, while yellow arrows indicate the goal direction in the scene. Section 5.2 says the robot first explored the junkyard and then underwent these evaluations without additional exploration. Compare whether the local path must bend around a new obstacle with whether the goal remains recognizable under appearance change. These are related but distinct demands on the learned controller and memory. e09e04e05
What it supports. The illustrated runs support the practical value of a goal representation that tolerates appearance changes while allowing local obstacle avoidance. Reusing the learned policy and explored environment under these changes is stronger evidence than simply replaying the original image sequence, but remains a qualitative demonstration in this particular setting.
Where the evidence stops. No per-condition success rates, trial counts or latent-invariance measurements accompany the figure. The claim that the bottleneck discards irrelevant appearance is the authors' interpretation, not a direct measurement shown by the trajectory overlays.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Visual goal discovery in unseen outdoor environments Eight unseen real-world environments, one target RGB image per environment; methods use the same 20-hour offline dataset and online fine-tuning. | RECON: 09:54 Exploration time (mm:ss), lower is better | ECR: 14:54; ViNG: 19:48; ANS: 21:00; PPO + RND: 21:18; InfoBot: 23:36. Table 1 favors RECON. These times imply a reader-calculated 33.6% reduction versus ECR, not a 50% time reduction. The caption says '50% faster.' Failure aggregation, trial counts and uncertainty are not specified; NTS is excluded after failing in every environment. e07e08 |
| Repeat navigation after goal discovery Second traversal to the discovered goal using prior exploration, in the Table 1 evaluation. | RECON: 00:26; SCT 0.92 Navigation time (mm:ss), lower; Success weighted by Completion Time (SCT), higher | ECR: 00:31 / 0.73; ViNG: 00:34 / 0.60; ANS: 00:45 / 0.33; PPO + RND: 00:47 / 0.22; InfoBot: 00:48 / 0.21. The memory-based system supports efficient return navigation. SCT is a dynamics-aware weighted metric, not an unweighted success percentage; the supplied paper cites its definition rather than giving its formula. e06e07e08 |
| Component ablations for discovery and recall Subset of unseen test environments; Table 2, distinct from Table 1. | Ours: 09:56 / 00:25.8 / 0.92 Exploration mm:ss / navigation mm:ss / SCT | Reactive: 11:54 / 00:37.4 / 0.63; Random Actions: 14:54 / 00:31.4 / 0.73; Vanilla Sampling: 14:06 / 00:28.7 / 0.83. Removing graph memory most degrades recall among these variants; random actions and removing the bottleneck also hurt. The subset size and uncertainty are absent, limiting causal precision. e11 |
| Undirected exploration without graph construction Five-minute exploration comparison in a novel environment; sampled-goal rollouts versus random action sequences. | Authors report 5× faster exploration with sampled goals. Exploration coverage/speed | Figure 6 shows broader trajectory spread for goal sampling than for random actions. This diagnostic removes graph construction, helping isolate exploration behavior. The figure lacks a quantitative coverage axis, and the text does not specify the coverage computation or uncertainty. e10 |
| Navigation under changed obstacles and appearance After exploring a junkyard, navigate toward the blue-dumpster goal without further exploration; test trash cans, cones, a car and different lighting/weather. | Figure 5 presents successful trajectories across the illustrated conditions. Qualitative goal-reaching trajectories | No numerical robustness baseline or success rate is supplied. The examples support robustness in this setting. They do not prove latent invariance or general reliability under arbitrary scene changes. e09 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 6. A graph-free diagnostic contrasts random-action wandering with sampled-goal rollouts. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The caption identifies the left panel as random actions and the right panel as exploration using the context-conditioned goal representation. In both panels, the yellow dot marks the start and the cyan trace records motion; green dashed outlines indicate the illustrated exploration frontiers. Keep the 20-meter scale in view when comparing the spatial spread. Section 5.3 describes five minutes of undirected exploration without building a topological graph. That matters: the comparison is intended to examine the exploratory behavior induced by a sampled goal, separately from graph-based frontier selection and later shortest-path recall. The panels themselves contain no quantitative coverage axis. e10e05
What it supports. Sampled-goal rollouts spread more broadly through the illustrated environment than random action sequences. The authors describe this as fivefold faster exploration. Because graph construction is excluded, the diagnostic supports the idea that the latent-goal controller contributes exploration structure before memory-based planning is added.
Where the evidence stops. The exact coverage computation, repeated-trial statistics and uncertainty behind the fivefold claim are absent from the supplied body. Do not estimate a new area ratio from the dashed outlines or treat this illustration as a goal-success measurement.
Table 2. The subset ablations distinguish exploration gains from the memory needed for efficient recall. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row with the variant definitions in Section 5.3. Reactive removes the topological graph. Random Actions retains frontier-based exploration but replaces goal-directed rollouts with random action sequences, matching the ECR variant. Vanilla Sampling learns goal-conditioned actions and distances without the information bottleneck. The Ours row reports the complete method. Compare discovery and subsequent navigation separately: a variant can still locate the target reasonably quickly while being worse at returning to it. These trials use a subset of unseen environments, so this table's 09:56 and 00:25.8 should remain attached to this protocol rather than replacing Table 1's 09:54 and 00:26. e11e07e08e13
What it supports. Removing memory changes recall time from 00:25.8 to 00:37.4 and SCT from 0.92 to 0.63. Random Actions takes 14:54 to explore versus 09:56 for the complete system, while Vanilla Sampling takes 14:06. Together the results support contributions from memory, structured exploration and compressed goals, with different effects on discovery and recall.
Where the evidence stops. The subset size and uncertainty are not reported, and the missing implementation appendix limits verification of how closely training and capacity were matched. The table supports component usefulness, not a unique causal explanation for visual invariance.
7. Analysis & limitations
7.1 What the evidence leaves open
The method does not explicitly account for value of information. The authors leave guarantees for efficient stochastic-policy and bottleneck exploration to future work. e12
Tables report point values without uncertainty or a clear treatment of unsuccessful trials. Baselines can fail on difficult goals, so timing comparisons alone do not establish an unconditional expected discovery time. e07e08e11
Visual robustness and the ablations are consistent with task-relevant compression, but do not uniquely establish that compression caused invariance. The full system also changes exploration behavior and adapts online. e04e05e09e11
7.2 Questions for discussion
- How well does prior density predict actual goal reachability under appearance shifts?
- Does the bottleneck improve exploration when online adaptation and graph construction are held fixed?
8. Reproducibility audit
8.1 Requirements and known gaps
Hardware described is a Clearpath Jackal with a forward-facing 170-degree RGB camera and onboard Jetson TX2. GPS and laser scanning serve safety during data collection; RECON uses only RGB. A reproduction needs matched control timing, action labels, distinct train/test environments and comparable online interaction budgets. e02e03e07
Recover the specified appendices before claiming faithful reproduction: Appendix A covers data/relabeling, B.1 graph and navigation subroutines, and B.2 architecture, hyperparameters and training. The supplied body does not establish numerical H, gamma, beta, epsilon or distance thresholds, optimizer settings, latent size, training compute or statistical aggregation rules. e04e05e07e13
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test whether prior density predicts executable goals
Reader-proposed check, not an experiment performed here: fix a trained model and choose observation–goal pairs spanning reachable routes, blocked routes and appearance changes. Record each goal latent's prior density and predicted temporal distance before executing matched, bounded navigation trials. Hold the controller horizon and safety interventions constant. Compare success and distance error across density bins, then sweep epsilon without retraining. Frequent high-density failures or no relationship between density and success would challenge the feasibility gate's interpretation. Numerical settings must be recovered or explicitly declared as new choices because the supplied appendix is absent. e04e05e09e13
Check 2: Factor bottleneck and memory under a matched exploration budget
Reader-proposed check, not a reproduction claim: run a two-by-two comparison with bottleneck on/off and topological memory on/off, using matched offline data, model capacity, online update counts and repeated starts. First compare five-minute graph-free exploration with a declared coverage measure, then compare discovery time, return time and SCT with failures included under a declared rule. Report variation across runs and environments. If memory improves return navigation but the bottleneck advantage disappears under these controls, that would narrow the mechanism supported by Table 2. Recover missing training and relabeling specifications before labeling the study faithful. e03e04e05e07e10e11e13
8.3 Reading coverage
Visual audit: All 11 supplied PDF pages were rendered and visually inspected, covering the title/authors, Figures 1–6, Tables 1–2, Equations 1–2, Algorithms 1–2, hardware/data/method/evaluation details and the reference-only ending. Every final crop was separately inspected; legends, relevant labels, table headings and scale markers are retained. Figure 3's conditioning arrows were cross-checked against Equations 1–2 and the text; its decoder-box latent-index discrepancy is disclosed. The six crops are Figures 2, 3, 5, 6 and Tables 1, 2. Referenced Appendices A and B are absent from this artifact, and no separate supplement or code is covered.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; 1 Introduction
- 2 Related Work
- 3 Problem Statement and System Overview; 3.1 Mobile Robot Platform; 3.2 Self-Supervised Data Collection & Labeling
- 4 RECON: A Method for Goal-Directed Exploration; 4.1 Learning to Represent Goals; 4.2 Goal-Directed Exploration with Topological Memory; 4.3 System Summary
- 5 Experimental Evaluation; 5.1 Goal-Directed Exploration in Novel Environments; 5.2 Exploring Non-Stationary Environments; 5.3 Dissecting RECON
- 6 Discussion
- Acknowledgments; References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- All four supplied text chunks and all 11 PDF pages were read. The complete supplied main paper is reviewed, but the Appendices A, B.1 and B.2 referenced on pages 3 and 5 are absent from this PDF, which ends with references. Separate supplemental material availability has not been fully verified.
- Text extraction does not reconstruct figure images; the retained PDF was therefore inspected for figures, equations, algorithms and table layout.
- The title and four authors match the catalog. The title-page footer identifies CoRL 2021, London, UK; the supplied catalog separately records publicationYear 2022 and a 2021 BibTeX year. No explicit revision number or date appears in the supplied paper; this report identifies the proceedings artifact by its supplied hash and does not establish differences from another edition.
- Code, the linked dataset and project website were not inspected. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and conference footer
Title matches Rapid Exploration for Open-World Navigation with Latent Goal Models. Authors are Dhruv Shah, Benjamin Eysenbach, Nicholas Rhinehart and Sergey Levine. Affiliations are UC Berkeley and Carnegie Mellon University. Footer identifies CoRL 2021, London, UK.
Go to primary source ↓e02PDF p. 3, Sections 3 and 3.1
Goal-image navigation uses partial RGB observations and continuous linear/angular velocities, without localization or known dynamics. Platform is Clearpath Jackal, with 170-degree RGB camera and Jetson TX2. GPS and laser scanner are safety controllers during data collection.
Go to primary source ↓e03PDF p. 2, Section 2, distance/policy paragraph; p. 3, Section 3.2; p. 5, Section 4.3(A)
Distance uses supervised regression and control uses goal-conditioned behavior cloning. Time-correlated random walks collected over 18 months provide prior data; time steps proxy distances, with relabeling details referred to Appendix A.
Go to primary source ↓e04PDF pp. 3–4, Section 4.1, Equations (1)–(2), Figure 3 and surrounding text
The current-view-conditioned goal encoder feeds a latent to the action/distance decoder, also conditioned on current view. Information bottleneck objective balances predictiveness and compression; variational training uses likelihood and beta-weighted KL to a standard normal prior. Figure 3 arrows agree with this conditioning; its decoder box prints the latent indices as z_g^t whereas Equation 2 and the node use z_t^g.
Go to primary source ↓e05PDF p. 4, Section 4.2 and Algorithm 1, lines 4–15; p. 5, cases (i)–(iii) and SubgoalNavigate paragraph
Goal prior density above epsilon selects the goal; otherwise proximity to the least-explored nearby node decides between sampling a prior latent and going to that node. Rollouts hold a latent fixed for H steps with changing observations, update the graph/data and fine-tune encoder and decoder for gamma epochs. Frontier selection uses counts and distance thresholds delta_1 and delta_2.
Go to primary source ↓e06PDF p. 5, Algorithm 2 and Section 4.3(B–C)
After exploration, associate current and goal observations with graph vertices, compute a shortest path, encode successive vertex images and execute subgoal navigation. Exploration includes online model fine-tuning.
Go to primary source ↓e07PDF pp. 6–7, Section 5.1, baseline descriptions and evaluation protocol
Five tabulated baselines share 20 hours of offline data with RECON and are fine-tuned online. Evaluation uses eight unseen environments and one target image per environment, measuring discovery and second-traversal time plus SCT. NTS metrics are excluded because it fails all environments. Trial counts, uncertainty, SCT formula and failure-time aggregation are not provided here.
Go to primary source ↓e08PDF p. 6, Table 1, all rows and caption; p. 7, Section 5.1 results discussion
Exploration/navigation/SCT: PPO+RND 21:18/00:47/0.22; InfoBot 23:36/00:48/0.21; ANS 21:00/00:45/0.33; ViNG 19:48/00:34/0.60; ECR 14:54/00:31/0.73; RECON 09:54/00:26/0.92. Caption calls RECON 50% faster than ECR. The text reports discovery of goals up to 80 m away in under 20 minutes.
Go to primary source ↓e09PDF p. 7, Figure 5 and Section 5.2; p. 8, opening continuation of Section 5.2
After junkyard exploration, goal reaching is evaluated without more exploration under trash cans, traffic cones, a car and lighting/weather changes. Figure panels label sunny, rainy/overcast and twilight. Successful illustrated paths motivate the authors' invariance interpretation; no numerical robustness table is supplied.
Go to primary source ↓e10PDF p. 8, Figure 6, caption and Section 5.3, first paragraph
A five-minute undirected exploration comparison without graph construction contrasts random actions with sampled-goal rollouts. Authors report 5× faster exploration. Figure shows trajectories, start markers, exploration-frontier outlines and a 20 m scale, but no quantitative coverage plot or uncertainty.
Go to primary source ↓e11PDF p. 8, Table 2, all rows; Section 5.3, variant definitions and discussion
Subset evaluation reports exploration/navigation/SCT: Reactive 11:54/00:37.4/0.63; Random Actions 14:54/00:31.4/0.73; Vanilla Sampling 14:06/00:28.7/0.83; Ours 09:56/00:25.8/0.92. Reactive removes graph memory, Random Actions replaces sampled-goal rollouts at the frontier, and Vanilla Sampling removes the information bottleneck.
Go to primary source ↓e12PDF p. 8, Section 6, final paragraph
Authors identify lack of explicit value-of-information reasoning and leave theoretical exploration guarantees to future work.
Go to primary source ↓e13PDF p. 3, Section 3.2; p. 5, Section 4.2 final paragraph and Section 4.3(A); pp. 9–11, References
The body refers data/relabeling to Appendix A, subroutine pseudocode to B.1 and architecture/hyperparameters/training to B.2. The supplied PDF instead ends with acknowledgments and references on pages 9–11 and contains none of these appendices.
Go to primary source ↓e14PDF p. 2, Figure 2 and caption
Goal-image input, exploration trajectory, observation-map nodes and a discovered route are shown alongside distinct training and unseen test environments. The aerial scene illustrates the run; the method operates from onboard visual observations.
Go to primary source ↓e15PDF p. 6, Figure 4 and caption; p. 7, final paragraph of Section 5.1
Illustrated recalled routes show RECON and ECR reaching the checkerboard target while other shown baselines collide; RECON follows the shorter route in that example. This is a selected qualitative trajectory comparison.
Go to primary source ↓8.5 Primary sources
Rapid Exploration for Open-World Navigation with Latent Goal Models ↗
PDF · 6,677 extracted words
Source fingerprint
e82e8711db8b565a4a6f67fed91ecce543f7b57cefdcc848fc124a6c2b70c2e7