Source-linked AI summary
EARS-Info: Information-Aware Event-Driven Reallocation for Multi-Robot Exploration
Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya
TL;DR
Multi-robot exploration must seek new information while limiting redundant sensing, congestion, and avoidable movement. EARS-Info adds information-aware relocation, return-aware routing, sensing reservations, and reusable graph fields. It achieves the highest mean operational score in every reported evaluation panel, with a 5.43% score improvement over EARS-Safe at the primary setting.
Problem
Multi-robot exploration must direct robots toward new information while limiting overlapping sensing, congestion, avoidable movement, and coordination overhead.
Source: §1 · p. 1Method
EARS-Info combines marginal sensing gain, feasible relocation, return-aware routing, sensing reservations, and reusable topology-dependent graph fields.
Source: §1 · p. 1Results
EARS-Info-v9 achieves the highest mean operational score in every reported panel; at the primary setting, score improves by 5.43% over EARS-Safe.
Source: §10 · p. 10Takeaways & Limitations
The evidence supports the complete-controller improvement, while component-level superiority remains outside the established evidence.
Source: §6.4 · p. 8, §10 · p. 10Generalization is limited by previously used evaluation cities, one training replicate, coarse maps, and unequal model capacity and training resources.
Source: §8 · p. 10
Abstract
from arXiv · showhide
We introduce EARS-Info, an information-aware extension of Event-Driven Ant Reallocation Search for multi-robot exploration on OpenStreetMap-derived grids. It combines marginal sensing gain, feasible relocation, return-aware routing, sensing reservations, and cached shortest paths. Across 168,000 episodes comparing 30 implementations on eight city maps, EARS-Info achieves the highest mean operational score in every evaluation panel. With eight robots and 3% communication dropout, it improves the score over EARS-Safe by 5.43%, increases weighted discovery from 88.88% to 98.46%, and reduces mean energy use by 2.62%. Score differences are positive in all eight cities ($p=0.0078125$; descriptive 95% city-bootstrap interval $[0.0159,0.0694]$), with no simulated incidents across 5,600 EARS-Info episodes. These results support the complete controller; individual component advantages remain uncertain. Generalization is limited by reused geography, one training replicate, coarse maps, and adapted comparison methods.
1 Introduction
EARS-Info addresses weaknesses in local scoring and relocation by making information gain, route feasibility, return travel, sensing overlap, and path reuse part of the decision process. The paper frames its main test as a comparison with EARS-Safe, while noting that the evaluation is exploratory on previously used geography.
Motivation
- Local scores can revisit sensed neighborhoods, relocation can duplicate sensing, and rebuilding shortest-path fields can dominate runtime.Event triggering alone does not ensure that a requested route is valuable or feasible.Source: §1 · p. 1
EARS-Info and its decision process
- EARS-Info scores newly sensed area, compares feasible distant gain with the best local opportunity, accounts for return travel, reserves sensing footprints, and reuses topology-dependent graph fields.Its policy reads observable uncertainty and visit information, not unsensed mission labels.Source: §1 · p. 1
Contributions
- The paper specifies EARS-Info within the EARS family, defines matched controls, compares 30 implementations on identical evaluation inputs, and reports city-level inference, rankings, task metrics, and artifact checks.The comparison comprises 168,000 episodes.Source: §1 · p. 1
Evaluation scope
- The main question is whether the complete information-aware controller improves on EARS-Safe; the evaluation is exploratory because the eight cities appeared in earlier development.The results are conditional on frozen models, not untouched geographic confirmation.Source: §1 · p. 1
2 Related Work
Prior work spans cooperative exploration, swarm allocation, MARL, robotic benchmarks, and safety-aware coverage. EARS-Info is positioned as an information-aware, event-driven extension evaluated in a static map-derived city abstraction.
2.1 Cooperative Exploration and Bio-Inspired Search
- Exploration research includes macro-action and asynchronous MARL, reinforcement-learning coordination, and frontier-led swarming; ant-based allocation also studies event triggers, learned intentions, and uncertain assignments.These studies motivate AntSwarmSafe as a serious algorithmic baseline.Source: §2.1 · pp. 1–2
- EARS asks whether global coordination can be applied sparsely when local search degrades, while adding information-aware route selection, caching, and persistent sensing reservations.Event-triggered mechanisms themselves have prior uses in allocation and bandwidth-aware MARL communication.Source: §2.1 · p. 2
2.2 MARL and Robotic Benchmarks
- Robotic benchmarks emphasize dynamics, embodiment, transfer, and reproducibility, while SafeSwarm-CityTwin focuses on urban map abstractions, runtime checks, interpretable swarm methods, and geographic comparisons.The cited discussion also connects MARL to persistent exploration and scalability challenges.Source: §2.2 · p. 2
2.3 Safety, Energy, Mapping, and Digital Twins
- Here, a city twin is a static map-derived abstraction of obstacles, restricted areas, bases, targets, uncertainty, and communication conditions, without live synchronization, calibrated vehicle dynamics, or physical field validation.Related work also addresses runtime filtering, energy-aware coverage, mapping, and digital-twin applications.Source: §2.3 · p. 2
3 Problem Formulation and Runtime Constraints
The study models mission exploration on a known grid with hidden mission locations, shared observations, and runtime checks. Its operational score combines task and operational metrics, while the implemented safeguards are not a formal safety guarantee.
3.1 Partially Observable Search With Shared State
- Robots move on a known grid, but mission locations and priorities remain hidden until sensed; this is mission exploration, not SLAM in unknown geometry.The model specifies traversable and obstacle cells, restricted cells, bases, robot positions, and batteries.Source: §3.1 · p. 2
- Policies directly access shared observation, uncertainty, pheromone, visit, position, battery, and communication fields; dropouts do not model bandwidth limits or delayed map exchange.Per-robot action scoring therefore does not imply a fully distributed implementation.Source: §3.1 · p. 2
3.2 Metrics and Operational Score
- Weighted discovery is the fraction of total mission priority weights found, while coverage, redundancy, distance, energy, and communication availability have distinct operational definitions.Communication availability measures online robot-time samples, not message efficiency.Source: §3.2 · p. 2
- The hardware-independent operational score combines bounded components, but early mission completion shortens rollouts, so energy, distance, coverage, and runtime are not equal-duration comparisons.Raw metrics accompany the score to expose tradeoffs.Source: §3.2 · p. 2
3.3 Runtime Safety Checks and Their Scope
- The runtime monitor checks boundaries, obstacles, restricted and occupied cells, battery-return estimates, and communication-loss counters, replacing invalid proposals when an admissible alternative exists.A proactive guard favors base progress before the hard reserve threshold.Source: §3.3 · p. 3
- The checks do not guarantee formal safe-set invariance: Manhattan distance can underestimate return routes, and fallback STAY may violate a rule.Zero incidents is empirical and limited to the implemented metric and tested scenarios; “Safe” does not mean certified deployment safety.Source: §3.3 · p. 3
4 Proposed Methods
The proposed methods develop from an observable-evidence local swarm rule to event-triggered relocation, then extend that process with marginal sensing gain, feasible return-aware routing, sensing reservations, and cached graph fields.
4.1 AntSwarmSafe
- AntSwarmSafe ranks candidate cells using observable novelty, uncertainty, sensed evidence, pheromone, congestion, and distance to an observable evidence proxy.It proposes the highest-scoring non-obstacle neighbor, including the current cell, before sequential runtime filtering.Source: §4.1 · p. 3
4.2 EARS-Safe: Event-Driven Ant Reallocation Search
- EARS-Safe triggers relocation on stagnation, repeated visits, or congestion, with checks beginning at step seven and relocation subject to cooldown and return behavior.Stagnation uses at least nine stored positions and a unique-cell ratio at most 0.46; eligible checks prioritize stagnation, revisit, then congestion.Source: §4.2 · pp. 3–4
- EARS-Safe allocates relocation goals using a global field and excludes obstacles and restricted cells, while graph-distance routing supports goal progress and return.Its paired comparison with AntSwarmSafe estimates the complete EARS implementation, not event triggering alone.Source: §4.2 · p. 3
4.3 EARS Family and Scope of Comparison
- The EARS family includes negative-pheromone and hierarchical MAPPO variants, while v9 controls separately disable relocation, alter request timing, or remove sensing reservations.These controls preserve the specified v9 checkpoint and routing configuration except for their named intervention.Source: §4.3 · p. 4
4.4 Marginal Sensing Gain and Local Action Selection
- EARS-Info-v9 combines marginal sensing gain, sensing-footprint reservations, feasible return-aware relocation, and cached graph distances.It is presented as a v9 information controller alongside matched diagnostic controls.Source: §4.4 · p. 4, §4.5 · p. 5, §4.6 · p. 5
4.5 Event Requests and Feasible Information Routing
- V9 requests relocation after stagnation or local information exhaustion, then accepts a feasible positive-gain goal only when its value clears the best local opportunity by at least 0.08.Accepted goals reserve sensing footprints; deadlines, retry backoff, cooldown, and return necessity govern route continuation or cancellation.Source: §4.5 · pp. 4–5
4.6 Caching and Computational Cost
- The graph-distance cache retains up to 96 goal fields and is invalidated when topology changes, reducing repeated graph work without guaranteeing constant-time planning.Its worst-case storage is O(96M), and each relocation request may still scan candidates and build a distance field.Source: §4.6 · p. 5
5 Experimental Methodology
The evaluation uses frozen OpenStreetMap-derived city grids and matched episodes across methods, with exploratory inference based on eight reused city geographies and one training replicate.
5.1 Geography, Observation, and Map Provenance
- The eight evaluation cities were excluded from current training and checkpoint selection, but had appeared in earlier development, making geographic evaluation exploratory rather than untouched confirmation.Maps are OpenStreetMap-derived 40×40 grids, and provenance checks establish consistency rather than fidelity to building footprints or physical mobility.Source: §5.1 · p. 5
- Figure 1 depicts observable sensing gain in both local exploration and relocation, with graph return and runtime action filtering remaining in execution.All methods use identical frozen map snapshots after provenance and consistency checks.Source: §5.1 · p. 5, Figure 1 · p. 6
5.2 Training, Selection, and Method Ownership
- The comparison includes 26 preceding and four v9 implementations, but one training replicate is available and several policy-style adaptations are not complete published-algorithm reproductions.The frozen implementations, rather than a general ranking of published MARL algorithms, are compared.Source: §5.2 · p. 6
5.3 Evaluation Grid and Paired Execution
- Each mission-stop and fixed-budget panel contains 76,800 episodes across eight cities, four team sizes, four dropout rates, and 30 methods.The primary setting uses eight robots and 3% dropout; shared maps, starts, episode seeds, and step-indexed communication draws support matched comparisons.Source: §5.3 · pp. 6–7
5.4 Primary Inference and Secondary Analyses
- Primary inference averages paired EARS-Info-v9 minus EARS-Safe score differences across eight cities and applies an exact two-sided sign-flip test.The city-bootstrap interval reflects between-city variation conditional on frozen checkpoints, not training-seed variability; secondary comparisons are exploratory and Holm-corrected.Source: §5.4 · p. 7
- The SWAP analysis changes hidden mission subsets on fixed geometry, but Chicago’s unchanged one-target view is excluded from changed-view inference.Chicago remains in the descriptive all-city panel, and other robustness analyses are post hoc descriptions using the same frozen episode data.Source: §5.4 · p. 7
6 Results
At the primary setting, EARS-Info-v9 outperforms EARS-Safe, and it has the highest mean score across the evaluation panels. The gain varies by city, while matched controls leave individual component advantages uncertain; runtime is lower and simulated incidents are absent in the tested episodes.
6.2 Performance Across Evaluation Panels
- EARS-Info-v9 has the highest equal-city mean score in the mission-stop, fixed-budget, all-city SWAP, and changed-view SWAP panels.The mission-stop and fixed-budget panels are kept separate, and Figure 2 shows score across team sizes and dropout settings for one fixed configuration.Source: §6.2 · p. 7, Table 4 · p. 8, Figure 2 · p. 9
6.3 Geographic Heterogeneity and Sparse-Map Sensitivity
- All eight city score contrasts are positive, ranging from +0.000373 in Chicago to +0.094265 in Rome, but their magnitudes vary substantially.Excluding Chicago gives a post hoc seven-city mean improvement of +0.047901; sparse maps and near-empty obstacle grids caution against treating city labels as equivalent geometric difficulty.Source: §6.3 · p. 7, Figure 3 · p. 9
6.4 Matched Controls and Limits of Mechanism Evidence
- Disabling relocation yields J = 0.618154 versus 0.814958 for full v9, but its adjusted secondary p-value is 1.0.Periodic relocation and the no-sensing-reservation variant also score below full v9, but neither comparison establishes an independent component advantage; the primary gain is a joint effect of local scoring, routing, return behavior, caching, and coordination.Source: §6.4 · pp. 7–8
6.1 Complete-Controller Improvement at the Primary Setting
- 5.43% relative: EARS-Info-v9 scores 0.814958 versus EARS-Safe’s 0.772998 at the primary setting, ranking first among 30 implementations.All eight city contrasts are positive; the exact sign-flip test gives p = 0.0078125 and the descriptive 95% city-bootstrap interval is [0.015908,0.069364].Source: §6.1 · p. 7, Table 3 · p. 8
6.5 Runtime and Simulated Incidents
- 0.419947 s per episode: EARS-Info-v9 is faster than EARS-Safe’s 3.089217 s, while EARS-Info-v9 records zero combined incidents across 5,600 evaluations.The runtime comparison is conditional on hardware, serial execution, and stopping rule; zero incidents establish simulated execution in tested episodes, not superior safety or physical safety.Source: §6.5 · pp. 8–9
6.6 Score Dependence
- EARS-Info-v9 remains above EARS-Safe and ranks first among all 30 implementations across 2,000 tested score-weight vectors.Its mean-score margin over EARS-Safe ranges from 0.03333 to 0.05073; these fixed-data sensitivities are not independent simulations or significance tests.Source: §6.6 · p. 9
7 Discussion
EARS-Info-v9 improves on EARS-Safe across the evaluated city abstractions, while its design offers a favorable measured balance of discovery, energy, and runtime in the stated benchmark. The improvement is attributable to the complete controller, not established as an independent benefit of individual components.
7.1 What Changes Relative to Earlier EARS
- Positive score differences across all eight evaluated abstractions support an improvement of the complete EARS-Info-v9 controller over EARS-Safe.The result does not independently identify a causal decomposition.Source: §7.1 · p. 9
- V9 values newly sensed cells, screens relocation using graph travel and return distances, coordinates marginal sensing, and reuses graph structure.The interpretation that sensing location matters beyond visited-cell count is consistent with the design, but is not an independently identified causal decomposition.Source: §7.1 · p. 9
7.2 Practical Interpretation
- Within the static, known-geometry benchmark, EARS-Info-v9 combines discovery, energy, and runtime favorably when shared uncertainty/visit fields and a traversable graph are available.The reported benefit does not directly transfer to unknown geometry, delayed map exchange, uncertain localization, continuous dynamics, or real battery discharge; those settings require additional observation/communication and execution models.Source: §7.2 · pp. 9–10
8 Limitations
The evaluation is exploratory and its generalization is constrained by reused geography, coarse maps, and limited training and comparison resources. Joint design changes and evaluation choices also limit component attribution and interpretation of some metrics.
8 Limitations
- Evaluation cities had appeared during development, making the study exploratory and limiting claims of independent geographic generalization.Coarse rasterization produces sparse maps, including Chicago’s single-target instance.Source: §8 · p. 10
- One training replicate and unequal model capacity and training resources constrain generalization and external MARL comparisons.Joint design changes prevent isolated causal attribution, and no secondary test survives multiplicity correction.Source: §8 · p. 10
- Shared state persists during dropout, execution safeguards differ, and early termination limits energy comparisons; runtime is platform-specific.Zero simulated incidents do not guarantee physical safety.Source: §8 · p. 10
9 Reproducibility and Artifact Availability
The frozen study package covers 168,000 episode records across 30 methods and provides materials for checking the reported analyses. Its documented runtime and reporting workflow also have specific reproducibility constraints.
9 Reproducibility and Artifact Availability
- The frozen eight-city-publication-v3 study contains 168,000 episode records covering 30 methods, maps, checkpoints, source snapshots, and training/runtime manifests.All 227 checked hashes matched, and independent recalculation reproduced scores, panel means, and the primary p-value.Source: §9 · p. 10
- The runtime manifest records software versions, hardware, and single-thread settings, but supplies no immutable image digest.Analysis scripts and an artifact inventory accompany the manuscript materials.Source: §9 · p. 10
- The CSV supports reported analyses, but the omitted SQLite journal and full diagnostics are required for the original journal-based reporting workflow.OpenStreetMap data retain separate attribution and licensing requirements.Source: §9 · p. 10
10 Conclusion
EARS-Info-v9 extends the EARS family and achieves the highest mean operational score across reported panels in the evaluated benchmark. The findings support a complete-controller improvement within the tested scope, while component superiority, independent geographic generalization, and physical safety remain unestablished.
10 Conclusion
- Across 168,000 matched evaluations of 30 implementations on eight OSM-derived city maps, EARS-Info-v9 achieves the highest mean operational score in every reported panel.At the primary setting, its score improves 5.43% over EARS-Safe, weighted discovery rises 9.58 percentage points, and mean rollout energy decreases 2.62%.Source: §10 · p. 10
- EARS-Info extends the EARS family with marginal sensing gain, graph-feasible event-driven relocation, sensing reservations, and reusable path fields.The findings support the complete-controller improvement in the tested benchmark.Source: §10 · p. 10
- The improvement is presented as interpretable and auditable, with explicit implementation, statistical, and environmental scope.Component-level superiority, independent geographic generalization, and physical safety remain outside the established evidence.Source: §10 · p. 10