acceptodds
Under review as a conference paper at ICLR 2027

RoboTopo: Inferring Hidden Communication Topology in LLM-Controlled Multi-Robot Systems

Abstract

Communication topology carries critical security information in Large Language Model (LLM) controlled multi-robot systems. It reveals how information flows across the system and which nodes or paths may serve as effective targets for attack. Despite its security relevance, it remains largely unexplored whether ordinary task interactions can reveal hidden communication topology and enable more effective subsequent attacks. In response, we introduce RoboTopo to infer the hidden communication topology, including the number of robots and planners and their directed connections, without prior knowledge of participants or access to internal messages. It combines physical execution constraints and controlled task dependencies with probability updates and adaptive probing. We construct the RoboCoord dataset with five multi-robot system configurations to evaluate RoboTopo. Across all runs, robot and planner counts are correctly recovered, with each identified node assigned an existence probability above 0.7. Retaining edges with predicted probabilities of at least 0.5 yields an average F1 score of 0.9917, compared with 0.9519 when requiring at least 0.7. Robustness tests under node or edge removal further identify all tested structural changes, with an average F1 score of 0.9767 for edge reconstruction. Using the inferred topology, we rank capability critical nodes and paths to guide attack propagation, increasing mean unsafe action coverage from 65.9% to 98.1% across three attack settings. Finally, we introduce StageGate, which restricts public reports to selected task checkpoints and reduces the overall topology reconstruction score by 47.2% while preserving progress feedback for most requests.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Embodied Intelligence, AI Security, Multi-Robot Systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6RXYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether external attackers can infer a multi-robot system’s communication topology from ordinary task requests and public feedback. RoboTopo achieves promising results on the authors’ RoboCoord dataset, and the inferred graphs help attackers spread unsafe actions more effectively. The authors also propose and evaluate a mitigation method. Overall, the paper reveals an important attack surface in multi-robot collaboration, though some methodological and experimental details need further explanation to better support its conclusions.

Critical strengths
  1. The paper shows that ordinary task feedback can reveal hidden communication structure through a single public interface, without access to internal messages or participant identities. This is an interesting security finding.

  2. The probes are well motivated by physical constraints. Simultaneous occupancy helps estimate robot counts, while matched task controls help separate execution failures from communication limitations.

  3. The attack experiments show that the inferred topology helps select more effective targets and spread unsafe behavior to more robots, demonstrating the practical security implications of topology leakage.

  4. StageGate provides a useful initial defense, reducing the reconstruction score while preserving progress feedback for most requests.

Critical weaknesses
  1. In the experiments verifying propagation attacks, the authors compared multiple models for InfectBot, External Entry Point Attack, and Privileged In-System Attack, whereas this paper analyzes only the performance of the GPT-4O model; the experimental results for other models are not yet known.

  2. In related work, the authors described topology detection approaches similar to digital multi-agent systems (such as MASLeak and CIA), but did not clearly explain why such methods are not applicable to embodied intelligence scenarios.

  3. Algorithms 1, 2, and 3 in the appendix are key manifestations of the method described in this paper; however, neither the main text nor the appendix provides an analysis of the method’s time complexity, space complexity, or probe cost. A reasonable analysis should be conducted in conjunction with actual probe scenarios.

Questions for authors
  1. The entire set of topology detection and subsequent attack propagation experiments is based solely on GPT-4o. Why? Please explain how different large language models would affect the effectiveness of topology inference and attack propagation (such as InfectBot, External Entry Point Attack, and Privileged In-System Attack, as mentioned by the authors). The authors should conduct cross-model experiments to demonstrate the method’s generality;

  2. For digital multi-agent topology inference methods such as MASLeak and CIA: If they can be adapted for use in this scenario, please provide comparison results showing the adapted versions; if they cannot be adapted, please explain in detail the underlying core differences and obstacles.

  3. Please analyze the temporal, spatial, and probe overheads of Algorithms 1–3 in the context of real-world detection scenarios, and include an explanation of how the algorithms handle boundary cases to improve the readability of the paper.

8 comments
AuthorsOpenReview ↗

Summary of Revisions for Reviewer 6RXY

Thank you for your review and support of our work. In response to your comments, we have made the following revisions to the manuscript.

  • Topology inference and propagation. Sections 4.1–4.3 and Tables 1, 2, and 4 cover GPT-4o, Gemini-2.5-Flash, and Qwen3. Table 4 reports all three propagation attack settings with the existing nine metrics.
  • MASLeak and CIA. Section 6 clarifies the differences in information sources and workflow models, including the adaptation needed for repeated robot execution and planner feedback.
  • Algorithms 1–3. Appendix D.7 adds time and space complexity, request costs, and handling of stopping conditions, missing reports, and undefined quantities.
AuthorsOpenReview ↗

[1/5] Response to Reviewer 6RXY: Models and participant recovery

To W1

  • Additional target models. To align with the model coverage of the propagation studies considered in our paper, we extended topology inference evaluation to Gemini-2.5-Flash and Qwen3 across all five graph configurations, with three runs per configuration. Their mean edge F1 scores are 0.9757 and 0.9961, respectively.

  • Attack propagation. Across nine model and attack settings with GPT-4o, Gemini-2.5-Flash, and Qwen3, RoboTopo achieves the highest mean NactN_{\mathrm{act}}, NdownN_{\mathrm{down}}, CC, and call AUC within each setting. Q1 reports all nine metrics from the main paper using the same definitions and aggregation rules.

To Q1: Topology inference

  1. Model selection and additional evaluation. We selected Gemini-2.5-Flash and Qwen3 to align our topology inference evaluation with the model coverage of the propagation-attack studies considered in our paper. InfectBot includes Gemini-2.5-Flash to represent practical robotic deployments that balance latency, efficiency, and reasoning capability. The communication-attack study includes the Qwen family to reflect model backends used in commercial robot systems. We evaluated both additional models across all five graph configurations, with three runs per configuration.

    • Robot and planner counts and edge F1.

      GraphN∗/K∗N^{\ast}/K^{\ast}GPT-4o E[N];N^\mathbb{E}[N];\hat NGPT-4o E[K];K^\mathbb{E}[K];\hat KGPT-4o mean edge F1Gemini E[N];N^\mathbb{E}[N];\hat NGemini E[K];K^\mathbb{E}[K];\hat KGemini mean edge F1Qwen E[N];N^\mathbb{E}[N];\hat NQwen E[K];K^\mathbb{E}[K];\hat KQwen mean edge F1
      G06/06.3750; 6/6/60.2500; 0/0/01.00006.3750; 6/6/60.2500; 0/0/01.00006.3750; 6/6/60.2500; 0/0/01.0000
      G17/07.3750; 7/7/70.2500; 0/0/00.95836.7917; 6/6/70.2500; 0/0/00.93877.3750; 7/7/70.2500; 0/0/00.9804
      G26/16.3750; 6/6/60.7500; 1/1/11.00006.3750; 6/6/60.7778; 1/1/10.96556.3750; 6/6/60.7500; 1/1/11.0000
      G34/14.3750; 4/4/40.8000; 1/1/11.00004.3750; 4/4/40.8000; 1/1/11.00004.3750; 4/4/40.8000; 1/1/11.0000
      G45/25.3750; 5/5/51.5000; 2/2/21.00005.3750; 5/5/51.4611; 2/2/20.97445.3750; 5/5/51.5000; 2/2/21.0000
      Σ\Sigma5.6000/0.80005.9750; 5.60000.7100; 0.80000.99175.8583; 5.46670.7078; 0.80000.97575.9750; 5.60000.7100; 0.80000.9961
    • Participant recovery and edge F1. Across the five graph configurations with three runs each, GPT-4o, Gemini-2.5-Flash, and Qwen3 recover the correct robot counts in 100% (15/15), 86.67% (13/15), and 100% (15/15) of runs, respectively. All three models recover the correct planner counts in 100% (15/15) of runs. Their mean edge F1 scores at a threshold of 0.5 are 0.9917, 0.9757, and 0.9961, respectively. Gemini’s robot-count errors are confined to G1, where it identifies six of seven robots in two runs; the third run recovers all seven robots and all nine true connections.

AuthorsOpenReview ↗

[2/5] Response to Reviewer 6RXY: Edge probabilities and coverage

To Q1 (continued): Topology inference

  1. Topology inference (continued).

    • Edge probabilities and coverage (Table 2 metrics).

      GraphEdgeGPT-4o pˉc+\bar p_c^+GPT-4o pˉm+\bar p_m^+GPT-4o CEC_EGemini pˉc+\bar p_c^+Gemini pˉm+\bar p_m^+Gemini CEC_EQwen pˉc+\bar p_c^+Qwen pˉm+\bar p_m^+Qwen CEC_E
      G0R→R0.97890.978990/900.95240.952490/900.97890.978990/90
      G1R→R0.94580.945825/270.96030.960324/270.98210.982126/27
      G2R→R0.98480.984830/300.97950.979527/300.98610.986130/30
      G2R→P0.72220.54179/90.72220.56179/90.72220.54179/9
      G2P→R0.75000.56259/90.74070.57569/90.75000.56259/9
      G3R→P0.75000.600012/120.75000.600012/120.71530.572212/12
      G3P→R0.80000.640012/120.80000.640012/120.76530.612212/12
      G4R→R0.98440.984418/180.98110.981118/180.98110.981118/18
      G4R→P0.70830.531312/120.70230.509611/120.70830.531312/12
      G4P→R0.70830.531312/120.70230.509611/120.70830.531312/12
      Σ\SigmaR→R0.97350.9735163/1650.96830.9683159/1650.98200.9820164/165
      Σ\SigmaR→P0.72690.557633/330.72480.557132/330.71530.548433/33
      Σ\SigmaP→R0.75280.577933/330.74770.575132/330.74120.568733/33
    • Connection recovery. Gemini and Qwen achieve robot-to-robot connection coverage of 96.36% (159/165) and 99.39% (164/165), respectively, compared with 98.79% (163/165) for GPT-4o. For both robot-to-planner and planner-to-robot connections, Gemini achieves 96.97% (32/33) coverage, while Qwen and GPT-4o achieve 100% (33/33). On supported true robot-to-robot connections, the mean conditional probabilities averaged across applicable graphs are 0.9683, 0.9820, and 0.9735 for Gemini, Qwen, and GPT-4o, respectively. These results show that topology recovery remains effective with both additional models, while performance varies across graph configurations.

Reviewer_VeaAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Summary: The main goal of the paper is to infer communication in a multi-robot system. Their attacker model is that the attacker can submit tasks and observe their outcome; the attacker does not have access to the communication between robots. The author identify three challenges they want to address: the number of robots and their task assignments are unknown, task failures may be caused by factors other than communication, and different coordination mechanisms may produce similar outcomes. The authors propose ROBOTO, by using tasks that require robots to occupy different locations simultaneously to estimate their number, comparing independent and dependent tasks to identify possible communication failures, and designing tasks that require different patterns of information sharing to infer the communication topology. The approach maintains probabilities over possible topologies and selects subsequent tasks to reduce uncertainty. To evaluate ROBOTO, the authors create a small benchmark with 5 topologies and show that their method works on this benchmark.

Critical strengths

Security is an important problem in multi-agent systems.

Critical weaknesses

My main issues with the paper relate to the setting of the problem, proposed solution, and evaluation.

Specifically:

-- the setting of the problem is a bit artificial, typical communication between robots is wireless, if the attacker is close enough to submit so many tasks, why it can not just monitor the network? The topologies they use are very simple, network monitoring would work very well. There is also a lot of work done in networking on finding out the network topology on systems like TOR, or enterprise networks, so monitoring will work also for more complex networks. The authors should provide a strong motivation of why this setting makes sense, why the attacker can not monitor the network.

-- In terms of the proposed solution it is not clear how well will it work when there are many robots with lots of dependencies between tasks. Also it seems to me that their approach of injecting so many tasks that are repetitive or slightly different, might make the attack easily detectable.

-- Evaluation is done on only 5 simple topologies with very few robots. It is not clear that these topologies are representative, and that the fact that their method works is not an artifact of the small size and simple setting of the network.

Questions for authors
  1. Why the attacker can not monitor the network?

  2. Why the simple topologies and small number of robots are representative for real systems

  3. Why the problem you are studying is important? What would an attacker need to know the network topology? Why for such a small network most likely wireless an attacker would need to so something so elaborate like your method, instead of just jamming or monitoring the entire network?

  4. Why with all the tasks you are injecting, a robot or a planner would not be suspicious and be able to deter your attack?

6 comments
AuthorsOpenReview ↗

Summary of Revisions for Reviewer VeaA

Thank you for the detailed feedback. In response to your comments, we have made the following revisions to the manuscript.

  • Practical motivation. Section 1 adds a public demonstration of robots coordinating tasks in response to user requests. Section 2.2 clarifies the distinction between transport connections and coordination decisions.
  • Security objective. Section 2.3 clarifies the attacker's observations and explains how the inferred topology guides target selection and attack order to influence physical task execution.
  • Costs. Appendix A.4 analyzes request budgets, and Appendix D.7 explains computational and request costs.
AuthorsOpenReview ↗

[1/3] Response to Reviewer VeaA: Threat model and network observations

To W1

  1. Our setting requires only access to the public task interface, without requiring visibility into internal communication traffic. Q1 explains this access distinction and why task feedback provides information beyond network connectivity.

  2. Our focus is the coordination structure among robot controllers and planners, and which nodes and paths are critical to completing physical tasks. This requires understanding robot capabilities and task dependencies in addition to communication connections, extending the inference target beyond wireless connectivity or network flow associations.

    We explain this distinction through system architecture, decision information flow, and task impact, drawing on industrial demonstrations, academic research, and our experimental results.

    • System architecture. Public requests followed by internal robot coordination already exist: Humanoid’s KinetIQ demonstration at GTC 2026 accepts voice requests, allocates tasks, and reports progress. Scalable (ICRA 2024) further studies centralized, decentralized, and hybrid LLM coordination, supporting the robot and planner roles considered here.

    • Decision information flow. Li et al. (Nature Communications, 2026) combine centralized TCP communication with decentralized task negotiation. Thus, a transport hub does not necessarily indicate a shared planner. Moreover, receiving a message does not establish the physical capability to execute the requested action. Our probes compare independent and dependent task execution to distinguish execution limitations from communication dependencies.

    • Task impact. We identify capability critical nodes and paths by their effect on feasible task participation. In G4, our selected path has a reference participation loss of 40%, compared with 0% for the Degree selection. Task capabilities and dependencies therefore matter beyond structural connectivity.

To Q1

Access through the public interface. Access to a public task interface does not imply visibility into internal message traffic. Our threat model considers an ordinary service user who can submit requests and observe public task outcomes, without access to exchanges among robot controllers and planners. We investigate whether this limited access already suffices to recover the hidden coordination structure. Even when network observations are available, they do not by themselves establish the robot capabilities and task dependencies needed to identify capability critical nodes and paths.

  • Why the task interface matters. Public task outcomes help infer which anonymous robots can fulfill the requested roles and whether prerequisite results can reach the robots performing subsequent subtasks. These associations allow us to assess how losing a node or path affects complete task assignments. For example, connectivity to a ground robot does not establish that it can replace a robot needed for an upper patrol.

  • Why existing network inference does not directly answer this question. SUMo (NDSS 2024) and DeepCoFFEA (IEEE S&P 2022) associate network flows, while NetCLR (CCS 2023) identifies visited websites. These outputs do not specify which robots can fulfill particular physical task roles or which exchanges are required for a complete task. Our analysis combines inferred coordination with execution capabilities and task dependencies to determine which nodes and paths matter for that task.

AuthorsOpenReview ↗

[2/3] Response to Reviewer VeaA: Scale and task interaction

To W2

  1. Team size and task dependencies. Our probing principles do not require prior knowledge of the robot count or coordination architecture. Simultaneous occupancy constrains the number of distinct robots, while independent execution controls help distinguish execution limitations from communication dependencies. More robots and task dependencies introduce more candidate connections and task assignments, which can increase computation and the number of requests needed. RoboTopo prioritizes requests by expected information gain, with controls and repetitions included in the same request budget. Our experiments cover chains, cycles, partial planner connectivity, and multiple planners, demonstrating recovery across these coordination patterns at the evaluated team sizes.

  2. Repeated instructions and small task revisions also occur in normal robot interaction.

    • Normal interaction patterns. LMPC (RSS 2024) reports an average of 2–3 user interaction rounds in successful sessions and 5–6 in unsuccessful sessions, as users refine instructions based on execution results. YAY Robot (RSS 2024) reports that 92 of 120 distinct commands in its bag packing intervention dataset appear at least three times across recorded task executions. These studies document repetition and refinement as part of legitimate robot use.

    • Industrial operation. Humanoid’s KinetIQ coordinates grocery picking, container handling, and packing. Its fleet orchestrator accepts new task requests, updates, and reassignments, while saved workflows can be executed again across the fleet. Repeated operations and revised task requirements are therefore supported features of these services.

    • Number of requests. RoboTopo achieves complete topology recovery with as few as 27 requests. Across all runs, the final robot and planner counts and edge F1 are reached with 34.2% fewer requests on average than the final stopping budgets. Subsequent probes examine remaining candidate coordination relationships and update their probabilities. The final budgets therefore include continued probing after these reconstruction results have already been reached.

To Q4

Repetition and small task changes alone do not distinguish malicious probing from legitimate task use. As discussed in W2, users also repeat and revise instructions to correct execution and update their requirements. RoboTopo requests permitted task execution, while the attacker infers topology by comparing the returned outcomes, without asking robots to disclose their internal configuration. Because the requested operations remain within the service’s permitted functions, the robot or planner can accept them as ordinary task instructions and updates.

Reviewer_ZBPHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper talks about a security problem in LLM-controlled multi-robot systems and their hidden communication topology on how an external user upon ordinary use of the system can use the information leaks to attack the robots. They propose RoboTopo, an attack that fetches the topology with using only normal Task requests and the public status updates their system returns without any internal access. This attack asks the robot to stay at several locations at once to know about the number of robots, detect the connections through tasks that depend on each other’s results, detect planners through reassignment requests and separate task failures from missing connections using the control requests. A Bayesian model updates the estimate after each response and picks the next most informative request. The setup is on Isaac Sim with Five setups with GPT-4o, the attack recovers robot and planner counts in all runs and reaches an edge F1 of 0.99. Then they use the leaked Topology to choose attack targets. Finally, they propose StageGate, a defense that limits which tasks statuses are reported that lowers the overall reconstruction score by 47.2% while keeping most progress feedback.

Critical strengths
  1. The paper identifies a new and practical way that robot teams leak their internal structure with just normal command requests unlike other works using software agents. They use the physical limits of the robots such as not being able to to be in two places at once and from whether the result can actually reach another robot.

  2. Running control requests before interpreting a failure avoids the common mistake of confusing a failed task with a missing connection. Including defense(StageGate) makes the work complete

Critical weaknesses
  1. Key design choices are never varied. Some results come from the model’s prior rather than the responses. The expected robot count is exactly 0.375 above the true count in every setup (Table 1), and the planner probability in G0 and G1 stays at 0.250 throughout (Figure 4 and 5). The attack also relies on accurate status updates that are supported through an added instruction (SP1) and all the experiments use only GPT4o with 4-7 robots. Testing these factors would show how much of the results come from the method itself.

  2. The runs shown take 44 to 76 requests (Figures 4 to 8), and many include unusual instructions such as asking robots to share findings ”directly…without asking anyone else to pass them on” (Task Probe 4). The paper does not show whether these probes would go unnoticed or how accuracy changes with fewer requests. This would make the practical threats more convincing.

Questions for authors

Clarification: Appendix H.1 says only the entry robot has the RESPONSE action, but every robot gives an external_response. Does the user see one reply per robot? If so, the robot count would leak directly without the occupancy probe?

  1. How do the counts and edge F1 change with a different prior, without the SP1 instruction, with a different LLM and with fewer requests?

  2. Since StageGate checkpoints follow a public catalog order (Section 5.1), can an attacker who knows this rule still recover the edges using small two-step requests?

7 comments
AuthorsOpenReview ↗

Summary of Revisions for Reviewer ZBPH

We appreciate your valuable feedback and support of our work. In response to your comments, we have made the following revisions to the manuscript.

  • Model coverage. Sections 4.1–4.2 and Tables 1–2 now report topology recovery with GPT-4o, Gemini-2.5-Flash, and Qwen3.
  • SP1 and request budgets. Appendix A.3 and Table 6 examine recovery with and without the SP1 confirmation clauses. Appendix A.4 and Figure 9 show how recovery changes with the number of public requests.
  • StageGate. Appendix G.2 and Table 10 examine requests containing two tasks under the known checkpoint rule.
  • Public feedback. Section 2.3 and Appendices C.4 and H.1 clarify that the public interface exposes the entry robot's reply and task status records, which do not identify the executing robots.
AuthorsOpenReview ↗

[1/5] Response to Reviewer ZBPH: Clarifications and priors

To W1

  1. Expected robot count. The constant 0.375 offset comes from the retained uncertainty about additional robots. For an identified count nn, the numerical count representation in the reported runs assigns probabilities of 0.75, 0.125, and 0.125 to nn, n+1n+1, and n+2n+2, respectively, giving E[N]=n+0.375\mathbb{E}[N]=n+0.375. The two additional candidate nodes have existence probabilities of 0.25 and 0.125, both below the recovery threshold of 0.5. Thus, this numerical convention explains the offset in the expectation without changing the recovered robot count N^=n\hat N=n.

  2. Planner prior. The 0.250 value in G0 and G1 is the retained prior of an unresolved planner candidate. In this sensitivity analysis, planner recovery additionally requires support from public task responses alongside the existence threshold. Under this criterion, the recovered robot and planner counts and edge F1 remain unchanged across planner priors of 0.25, 0.50, and 0.75 (see Q1).

  3. SP1 instructions. The confirmation clauses encourage robots to base completion reports on actual execution results. Without these clauses, execution prerequisite violations were more frequent, while count accuracy and mean edge F1 remained unchanged across all five graphs (see Q1).

  4. Team size. Our teams of four to seven robots are comparable in size to those evaluated in Scalable and COHERENT. Our probing principles do not require prior knowledge of team size, although more robots and task dependencies can increase computation and the number of requests needed. See our responses to Reviewer VeaA, W2 and Q2 for the scalability discussion and representative robot teams.

To W2

  1. Request budget. The 44–76 requests shown in the figures are final stopping budgets. Complete topology recovery was already observed at 27–53 requests in 14 of the 15 runs. The remaining run reached an edge F1 of 0.875. The table under Q1 shows how participant counts and edge F1 change with fewer requests.

  2. Task wording and detectability. The quoted instruction specifies a condition for task execution. Task Probe 4 requires inspection findings to be delivered directly before a dependent task starts; otherwise, that task remains waiting. The robot or planner can process this as an execution requirement within the permitted service functions. The request does not ask for internal identities, messages, or connections: the attacker infers connections by comparing public task outcomes. Repeated requests and revisions also occur during normal service use, as users refine tasks based on execution results. See also our response to Reviewer VeaA, Q4, which explains how repeated and revised requests can be interpreted as ordinary task instructions within the service’s permitted functions.

To Clarification

All robots share the external_response field, yet only the entry robot’s reply is exposed to the user under our threat model. This reply is accompanied by task status records. Since one robot can complete several tasks sequentially, counting these records does not reveal the robot count.

To Q1: Planner priors

  1. Different priors: planner existence. We varied the initial planner existence probability across 0.25, 0.50, and 0.75, applying the planner recovery criterion in W1 to the same public task responses. All three priors yield identical node and edge sets, with correct robot and planner counts in all 15 runs and a mean edge F1 of 0.9917.
AuthorsOpenReview ↗

[2/5] Response to Reviewer ZBPH: SP1 and request budgets

To Q1 (continued): SP1 and fewer requests

  1. Without SP1 confirmation clauses. The SP1 confirmation clauses are intended to keep task completion reports consistent with actual execution outcomes. We compared topology recovery with and without these clauses. Across five graphs and three runs per graph, both conditions recovered the correct robot and planner counts in all 15 runs, with the same mean edge F1 of 0.9917. After removal, physical action attempts violating execution prerequisites increased from 11/94 (11.70%) to 17/86 (19.77%) in G4; the corresponding rate across all graphs increased from 6.80% to 8.10%. These results suggest that the clauses can help agents follow execution requirements, while the main recovery results remain stable without them.

    ConfigurationSP1 clausesRobotsPlannersMean edge F1R → R: conditional / marginalR → P: conditional / marginalP → R: conditional / marginalCoverage: R → R; R → P; P → R
    G0Retained601.00000.9789 / 0.9789N/AN/A90/90; N/A; N/A
    G0Removed601.00000.9789 / 0.9789N/AN/A90/90; N/A; N/A
    G1Retained700.95830.9458 / 0.9458N/AN/A25/27; N/A; N/A
    G1Removed700.95830.9672 / 0.9672N/AN/A25/27; N/A; N/A
    G2Retained611.00000.9848 / 0.98480.7222 / 0.54170.7500 / 0.562530/30; 9/9; 9/9
    G2Removed611.00000.9848 / 0.98480.7222 / 0.54170.7500 / 0.562530/30; 9/9; 9/9
    G3Retained411.0000N/A0.7500 / 0.60000.8000 / 0.6400N/A; 12/12; 12/12
    G3Removed411.0000N/A0.7500 / 0.60000.8000 / 0.6400N/A; 12/12; 12/12
    G4Retained521.00000.9844 / 0.98440.7083 / 0.53130.7083 / 0.531318/18; 12/12; 12/12
    G4Removed521.00000.9789 / 0.97890.7083 / 0.53130.7083 / 0.531318/18; 12/12; 12/12
    OverallRetained——0.99170.9735 / 0.97350.7269 / 0.55760.7528 / 0.5779163/165; 33/33; 33/33
    OverallRemoved——0.99170.9774 / 0.97740.7269 / 0.55760.7528 / 0.5779163/165; 33/33; 33/33
  2. Fewer requests. In our topology reconstruction experiments, RoboTopo achieves complete recovery with as few as 27 requests. Across all runs, the final robot and planner counts and edge F1 are already attained with 34.2% fewer requests on average (27–47% across runs) than the final stopping budgets. The table also reports results at 23–24 requests, with ranges indicating minima and maxima across the three runs per configuration.

    ConfigurationRequests: earlier → final(Robots, planners): earlier → finalEdge F1: earlier → final
    G024 → 44–45(6, 0) → (6, 0)0.909 → 1.000
    G124 → 62–78(7, 0) → (7, 0)0.000–0.615 → 0.875–1.000
    G224 → 71–74(6, 0) → (6, 1)0.667–0.720 → 1.000
    G323 → 44(4, 1) → (4, 1)0.545 → 1.000
    G424 → 76(5, 0) → (5, 2)0.600 → 1.000

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.