acceptodds
Under review as a conference paper at ICLR 2027

DuoMind: Enabling Distributed Multi-Robot Coordination with Semantic Communication

Abstract

Vision–language models (VLMs) and vision–language–action models (VLAs) have recently driven rapid progress in general-purpose robotic agents, yet most progress has focused on single-robot settings. Extending these capabilities to multi-robot systems remains challenging because robots must coordinate long-horizon behaviors while maintaining reliable, fine-grained execution. We introduce DuoMind, a distributed hierarchical framework for multi-robot coordination through semantic communication. Each robot uses a VLA-based action model for low-level execution and a VLM-based orchestrator for high-level reasoning and inter-agent coordination. At each planning step, the orchestrator at each robot reasons over the task instruction, local observations, and messages received from other robots. It then generates low-level instructions for the action model and semantic messages for peer robots. This architecture exploits the complementary strengths of pretrained models by combining the semantic reasoning capabilities of VLMs with the precise action-generation capabilities of VLAs. To address the scarcity of benchmarks for multi-robot coordination, we further develop RoboPoly, a benchmark comprising long-horizon manipulation tasks that require coordinated, closed-loop execution under distributed control. Experiments on RoboPoly and RoboTwin demonstrate that DuoMind improves multi-robot task performance, while ablation studies confirm the contributions of hierarchical orchestration and semantic communication. More details are available on our project page.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotics, Reasoning, Multi-Agent System, Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9vFJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DuoMind is a distributed hierarchical framework for two-robot coordination. Each robot uses a VLM orchestrator for high-level decisions and message generation, together with a VLA policy for continuous low-level control. Robots exchange structured natural-language messages about intentions, subgoals, and beliefs. The paper also introduces RoboPoly, a seven-task manipulation benchmark, and evaluates on eight RoboTwin tasks adapted to distributed arm control. Evaluation uses 400 rollouts per task, direct VLA baselines, communication ablations, and an alternative action model. Results show improvements on several tasks, particularly Cook Pot, where communication increases success from 1.00% to 39.25% within the hierarchical system.

Critical strengths
  1. The framework connects language-based coordination to closed-loop continuous manipulation rather than evaluating high-level plans alone. Tasks such as handovers and jointly lifting a pot require coordination to survive execution, making the system and benchmark relevant to robot foundation-model research.

  2. The communication ablation supplies concrete evidence that messages can matter within the proposed system. The large Cook Pot improvement is particularly consequential because it corresponds to a task with a clear synchronization requirement, rather than relying only on qualitative examples of cooperation.

Critical weaknesses
  1. The main comparison confounds high-level orchestration with predefined task structure and supervision. The prompt on page 20 supplies arm-specific candidate stages and their order, requires sequential advancement, and forbids skipping or returning to earlier stages. On RoboPoly, DuoMind also uses stage-level training instructions while direct VLA baselines use whole-task instructions. Consequently, the results do not establish how much improvement comes from VLM reasoning rather than the supplied stage program, finer supervision, or progress rules.

  2. The specific contribution of semantic communication is not established beyond access to peer information. A no-message ablation demonstrates that information sharing helps, but not that rich natural-language messages outperform simple stage and completion signals. Since the orchestrators also share a global camera, the benefit should be distinguished from compensating for otherwise invisible physical state. This distinction is central to the paper's positioning around distributed semantic coordination.

Questions for authors
  1. Clarification of the controller and comparison. Please provide the complete candidate-stage sequences and the precise forced-advancement rule. Is there a comparison using the same low-level weights, stage supervision, and ordered candidates with a simple visual stage controller? If not, which evidence separates autonomous orchestration from execution of the supplied stage program?

  2. Clarification of communication and observations. Which inputs are private to each orchestrator, and which global or wrist-camera observations are supplied to the VLM and VLA separately? Do existing results show that intention/subgoal/belief messages add value beyond compact stage identifiers and completion signals?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9vFJ (1/3)

We sincerely thank Reviewer 9vFJ for their constructive feedback and valuable suggestions. Below, we address the reviewer's comments and provide additional clarifications on the key concerns raised.

Weakness 1

On RoboPoly, DuoMind also uses stage-level training instructions while direct VLA baselines use whole-task instructions. Consequently, the results do not establish how much improvement comes from VLM reasoning rather than the supplied stage program, finer supervision, or progress rules.

Thank you for your insight on this important point. On one hand, we adopt an orchestrator with candidates to address the challenges of visual grounding and reasoning in complex multi-robot tasks, particularly given the limitations of the compact VLM used in our original experiments. Candidate-based selection is also a common approach to structuring VLM-based robotic decision-making [1][2], as low-level action models may struggle to reliably execute arbitrary instructions beyond their training distribution. On the other hand, recent advances in foundation models, such as GPT-6 Astra, have introduced stronger visual and embodied reasoning capabilities, potentially reducing the need for predefined task structures. To investigate this, we add experiments using GPT-6 Astra as the orchestrator, without any candidate lists and sequential advancement, and compare against the constrained setting under matched execution budgets. The experiment results shown as below:

Table 1. Comparison with GPT-6 orchestrator. Success rates (%) across three RoboPoly tasks.

MethodCook PotPrepare SnackClean Table
DuoMind w/ π0.5\boldsymbol{\pi}_{0.5}39.2518.0024.75
DuoMind GPT-6 as orchestrator12.5010.0020.00
π0.5\boldsymbol{\pi}_{0.5} only1.0016.2516.50

We observe that GPT-6 Astra performs below the original Qwen3-VL-4B orchestrator, while still surpassing the π0.5\pi_{0.5}-only baseline. Without candidate constraints, GPT-6 Astra produces more diverse instructions, some of which deviate from the action model's training distribution and are difficult to execute reliably. This indicates that the performance difference stem partly from the action model's limited ability to follow unconstrained instructions, rather than from the orchestrator's reasoning capability alone.

Meanwhile, the improvement over the VLA-only baseline provides further evidence of the effectiveness of DuoMind's distributed hierarchical coordination. These findings also emphasize that overall performance depends on the compatibility between high-level reasoning and low-level execution. A stronger foundation model does not necessarily yield better results when its instructions are not well aligned with the action model's capabilities. Nevertheless, DuoMind's modular design allows both components to be improved independently, suggesting that integrating more capable orchestrators with better-aligned action models could further enhance multi-robot coordination.

[1] Nasiriany et al. PIVOT: Iterative Visual Prompting Elicits Actionable Knowledge for VLMs. Secs. 3.1–3.2. Paper

[2] Buoso et al. Select2Plan: Training-Free ICL-Based Planning through VQA and Memory Retrieval. IEEE RA-L, 10(11), 11267–11274. Sec. III. Paper

AuthorsOpenReview ↗

Response to Reviewer 9vFJ (2/3)

Weakness 2

The specific contribution of semantic communication is not established beyond access to peer information. A no-message ablation demonstrates that information sharing helps, but not that rich natural-language messages outperform simple stage and completion signals. Since the orchestrators also share a global camera, the benefit should be distinguished from compensating for otherwise invisible physical state. This distinction is central to the paper’s positioning around distributed semantic coordination.

Thank you for this insightful comment on semantic communication. We acknowledge that a short signal can be sufficient for simple synchronization movements. However, DuoMind targets coordination across complex, long-horizon tasks, where each robot requires richer shared information to reason about its peers’ near-term intentions and expected outcomes. This motivates our inclusion of intention, subgoal, and belief fields in the communication messages. To better quantify the contribution of each component, we add ablation experiments that individually remove the intention, subgoal, and belief fields.

Table 2. Ablation study on communication components. Success rates (%) across RoboPoly tasks.

MethodCook PotPrepare SnackClean Table
DuoMind w/ π0.5\boldsymbol{\pi}_{0.5}39.2518.0024.75
No intention0.0018.0019.00
No subgoal0.0016.0013.00
No belief0.0019.0020.00
No communication0.0013.5017.25

The results show that removing individual message components can lead to performance degradation, as agents lose important information about their peers' intentions, goals, or beliefs needed for effective coordination. This is particularly evident in Cook Pot, where successful execution requires precise synchronization between the two robots. This further highlights the importance of our message component design in enabling effective multi-robot coordination.

We apologize for the confusion caused by the term "global camera," which may misleadingly suggest access to complete global information. In fact, this camera is positioned at the side of the workspace and provides only a limited side view of the environment, rather than full observability of the entire workspace or access to privileged physical states. To avoid this misunderstanding, we have renamed it the "shared camera" throughout the paper.

This camera configuration is fully compatible with our distributed framework, as it provides a common visual observation of the shared workspace without compromising the decentralized nature of decision-making. While the shared camera allows agents to partially observe their peers' behavior, it does not directly reveal their internal intentions, subgoals, or beliefs, which are explicitly conveyed through inter-agent semantic communication. Importantly, our no-message ablation retains identical camera inputs, isolating the benefit of explicit communication from visual access. Furthermore, the additional communication ablations described above quantify the contributions of individual message components, providing a more detailed assessment of the effectiveness of semantic communication in distributed coordination.

AuthorsOpenReview ↗

Response to Reviewer 9vFJ (3/3)

Question 1

Clarification of the controller and comparison. Please provide the complete candidate-stage sequences and the precise forced-advancement rule. Is there a comparison using the same low-level weights, stage supervision, and ordered candidates with a simple visual stage controller? If not, which evidence separates autonomous orchestration from execution of the supplied stage program?

As detailed in Appendix I, we provide the orchestrator with the following system prompt to enable candidate-based instruction selection:

This ManiSkill task supplies an arm-specific candidate set. The L1/R1-style prefixes are metadata tags, not part of the low-level instruction. Select exactly one candidate, but put only the sentence after its colon in low_level_instruction; do not invent, paraphrase, include the tag, or select another arm's instruction.

Permitted candidates for this <ARM_SIDE> arm:

<THIS_ARM_CANDIDATES>

Here, <ARM_SIDE> and <THIS_ARM_CANDIDATES> are predefined in the task configuration. For example, in Food Serve, the candidate instructions are:

L1: pick up mug and place on tray

L2: pick up bread and place beside mug on tray

R1: pick up bread and place in the middle

We also specify the following stage-advancement rule as provided in Appendix I:

Progress rule: do not keep issuing a subgoal that is already complete. Review the current image and your own history when available. If visual evidence or your own prior belief says the current subgoal's observable result is complete, you MUST select a different permitted candidate that advances the unfinished high-level task. Otherwise repeat the current candidate. Do not advance merely because time passed, and visually verify claims from the other agent.

In addition to these prompt-level instructions, we introduce a stagnation-prevention mechanism that triggers instruction advancement when the orchestrator repeatedly selects the same instruction due to the fault from VLM's limited visual capability. For example, we set the threshold to 6 high-level planning steps for Food Serve and 10 steps for Prepare Snack, preventing prolonged stagnation at a single stage.

Regarding orchestration, we adopt candidate-assisted instruction selection to address the challenges of visual grounding and reasoning in complex multi-robot tasks, particularly given the limitations of the compact VLM used in our original experiments. Candidate-based selection is also a common approach to structuring VLM-based robotic decision-making. Furthermore, our GPT-6 Astra experiments discussed in Weakness 1 demonstrate that a stronger reasoning backbone can operate without predefined candidates or stage-progression constraints, further validating the effectiveness of our framework when equipped with a more capable orchestrator. We emphasize that our primary contribution lies in the hierarchical framework for distributed multi-robot coordination, while the performance reported in our original experiments is partly constrained by the reasoning capabilities of the VLM available at the time.

Question 2

Clarification of communication and observations. Which inputs are private to each orchestrator, and which global or wrist-camera observations are supplied to the VLM and VLA separately? Do existing results show that intention/subgoal/belief messages add value beyond compact stage identifiers and completion signals?

Thank you for the opportunity to clarify the observation and communication setup. We apologize again for the misleading term "global camera," which we have now renamed "shared camera" throughout the paper. This camera is positioned at the side of the workspace and provides only a limited side view, rather than complete global information. Both the orchestrator (VLM) and the action model (VLA) receive observations from this shared camera and the robot's own wrist camera, which remains private to each robot. Importantly, DuoMind does not require a shared camera, as each robot can independently obtain visual observations of the workspace. In our experiments, we use a shared camera to reduce the number of cameras and computational overhead. Therefore, this observation configuration is fully compatible with our distributed multi-robot framework.

The ablation experiment for individual components in the shared messages are presented in our response to Weakness 2.

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

DuoMind pairs a VLM orchestrator with a VLA controller for each robot. The orchestrators select language instructions and exchange structured semantic messages to coordinate their actions. The paper introduces RoboPoly and evaluates the framework on RoboPoly and selected RoboTwin tasks. The simulation results show gains over the evaluated baselines, with communication ablations indicating a substantial benefit on tasks that require synchronized execution.

Critical strengths
  1. The language interface separates high-level coordination from low-level control and allows each robot to retain its own VLA controller. This makes the division of responsibility between the two levels easy to follow.

  2. The communication ablation provides direct evidence that the robots benefit from exchanging information, particularly on tasks requiring synchronized actions.

Critical weaknesses
  1. All results are in simulation. This is a significant gap for a method whose main contribution is closed-loop coordination between independently controlled robots. Physical execution errors, unequal action durations, inference latency, and asynchronous messages can change whether coordination prevents a collision or deadlock. The paper needs a real-robot experiment on synchronized lifting or handover, reporting success, timing mismatch, and failure modes.

  2. The recent Astra embodied-policy study [1] raises a relevant baseline question: how does a GPT-6 Astra orchestrator perform with the same VLA controllers, admissible instructions, observations, and call budget? Running it with and without the full DuoMind communication protocol would test the contribution of the protocol under a stronger backbone. Direct or Hybrid control is another useful reference if a compatible action adapter is available.

[1] Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies. arXiv:2609.38537. https://arxiv.org/abs/2609.38537

  1. On RoboPoly, DuoMind receives low-level instruction supervision, while the VLA-only baseline is trained with high-level task instructions. The comparison changes instruction granularity, hierarchy, and communication together. A controller trained with the same low-level labels but using fixed or state-based scheduling is needed to separate these effects under matched data and execution budgets. The no-communication ablation does not resolve the supervision difference.

  2. The no-communication ablation shows that exchanging information helps, but not that the full semantic message is necessary. How much of the gain remains with ready/wait or stage-status signals? Ablating intention, subgoal, and belief fields would test the value of the richer content. Message frequency, communication and inference overhead, and sensitivity to delayed messages also matter for judging whether the protocol works under realistic timing.

Questions for authors

Please see the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Aadf (1/2)

We sincerely thank Reviewer Aadf for their constructive feedback and valuable suggestions. Below, we address the reviewer's comments and provide additional clarifications on the key concerns raised.

Weakness 1

The paper needs a real-robot experiment on synchronized lifting or handover, reporting success, timing mismatch, and failure modes.

Thank you for this thoughtful suggestion on physical evaluation. Real-world multi-robot experiments entail substantial resources. A real-world multi-robot study requires at least two robots in a shared workspace, coordinated demonstration collection that can involve simultaneous teleoperation by multiple operators, and additional computation for agent fine-tuning. Given these practical constraints, the present study focuses on simulation to explore and get an in-depth understanding first. Our hope is to stimulate interest in this important area in upcoming multi-robot systems. It is nontrivial to develop the two-system framework to distributed multi-robot coordination, providing a foundation for subsequent real-world studies. Our work establishes a the two-system framework to distributed multi-robot coordination, providing a solid basis for this challenging and open field. We agree with reviewer's insight and we will build the experimental platform to validate multi-robot coordination through real-robot experiments in the future work.

Weakness 2

The recent Astra embodied-policy study raises a relevant baseline question: how does a GPT-6 Astra orchestrator perform with the same VLA controllers, admissible instructions, observations, and call budget? Running it with and without the full DuoMind communication protocol would test the contribution of the protocol under a stronger backbone.

Thank you for this insightful suggestion. We agree that comparing against newly emerging approaches is valuable, especially given the promising progress in agentic robotics. We have also been closely following recent developments along this line, which came out after we finished this work. During this discussion phase, we conducted additional experiments using GPT-6 Astra as the orchestrator, removing the predefined instruction candidate constraints to allow the orchestrator to generate instructions from a more diverse instruction space. The results are presented below:

Table: Comparison with GPT-6 orchestrator. Success rates (%) across three RoboPoly tasks.

MethodCook PotPrepare SnackClean Table
DuoMind w/ π₀.₅39.2518.0024.75
DuoMind GPT-6 as orchestrator12.5010.0020.00
π₀.₅ only1.0016.2516.50

We observe that GPT-6 Astra achieves lower performance than the original Qwen3-VL-4B orchestrator, but still outperforms the π0.5\pi_{0.5}-only baseline. On the one hand, removing the candidate constraints allows GPT-6 Astra to generate more diverse instructions, some of which fall outside the training distribution of the action model and cannot be reliably executed. This suggests that the performance gap is partly attributable to the limited instruction-following capability of the action model, rather than solely to the orchestrator's reasoning capability.

On the other hand, the improvement over the VLA-only baseline further supports the effectiveness of DuoMind's distributed hierarchical coordination framework. Importantly, overall performance depends on both the orchestrator and action model, and their capabilities must be well aligned. Simply replacing the orchestrator with a stronger foundation model does not necessarily improve performance, particularly in complex multi-robot coordination tasks. Nevertheless, these results highlight the flexibility and potential of DuoMind. We expect that combining stronger orchestrators, such as GPT-6 Astra, with better-aligned action models could further improve coordination performance in future work.

AuthorsOpenReview ↗

Response to Reviewer Aadf (2/2)

Weakness 3

A controller trained with the same low-level labels but using fixed or state-based scheduling is needed to separate these effects under matched data and execution budgets.

Thank you for raising this important point. We add a control experiment using a VLA trained with the same low-level instructions and a state-based oracle orchestrator for high-level planning. The oracle uses simulator task states to determine stage completion and select the next admissible instruction, while keeping the training data and execution budgets fixed. This comparison shown below controls for instruction granularity and provides a reference with privileged state access for evaluating the scheduling component.

Table: Comparison with state oracle. Success rates (%) across three RoboPoly tasks.

MethodCook PotPrepare SnackClean Table
State Oracle47.0020.0032.00
DuoMind w/ π₀.₅39.2518.0024.75
π₀.₅ only1.0016.2516.50

The oracle experiment demonstrates how effectively the action model can follow correct instructions from an accurate planner to complete the task. As expected, the oracle experiment with a "perfect" orchestrator achieves a higher success rate than DuoMind with a VLM-based orchestrator. Comparing the oracle, DuoMind, and π0.5\pi_{0.5}-only results demonstrates that our hierarchical design with a VLM-based orchestrator improves multi-robot coordination over the VLA-only baseline, while the oracle's stronger performance highlights the potential for higher overall performance by improving the orchestrator's reasoning capabilities.

Weakness 4

Ablating intention, subgoal, and belief fields would test the value of the richer content. Message frequency, communication and inference overhead, and sensitivity to delayed messages also matter for judging whether the protocol works under realistic timing.

Thank you for these helpful suggestions on the communication analysis. We add controlled ablations that individually remove the intention, subgoal, and belief fields to quantify their contributions to coordination. We also evaluate the impact of reducing the communication frequency by half and introducing a 300 ms communication delay to assess robustness to communication frequency and latency.

Table: Ablation study on communication components and robustness. Success rates (%) across RoboPoly tasks.

MethodCook PotPrepare SnackClean Table
DuoMind w/ π₀.₅39.2518.0024.75
No intention0.0018.0019.00
No subgoal0.0016.0013.00
No belief0.0019.0020.00
Communication delay0.0017.0016.00
Low communication freq.0.0015.0016.00
No communication0.0013.5017.25

The results show that removing individual message components can lead to performance degradation, as agents lose important information about their peers' intentions, goals, or beliefs needed for effective coordination. This is particularly evident in Cook Pot, where successful execution requires precise synchronization between the two robots. This further highlights the importance of our message component design in enabling effective multi-robot coordination.

When reducing the communication frequency by half or introducing a 300 ms delay, we observe some performance degradation, but DuoMind maintains reasonable robustness, with several tasks still achieving successful coordination under these challenging conditions. In practical multi-robot coordination, communication typically occurs over short distances, where latency is expected to be considerably lower. Moreover, our compact semantic messages introduce limited communication overhead, allowing more frequent information exchange. These results demonstrate DuoMind's robustness to communication constraints under realistic deployment conditions.

Dear Reviewer Aadf:

As the author–reviewer discussion period draws to a close today, we would appreciate any further thoughts you may have on whether our response addresses your main concerns.

As noted in our earlier response, our hope is to take initiatives to study multi-robot systems (which has barely been explored) and stimulate interest in this important area. It is nontrivial to develop the proposed two-system framework for distributed multi-robot coordination - it took us more than an year to work out this paper. We hope this study opens a door to this very important area in physical AI. Provided that you agree with this, we will appreciate you can improve your rating.

Reviewer_Lf8ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the DuoMind distributed hierarchical multi-robot collaborative framework, which utilizes VLM for high-level scheduling and structured semantic communication, and VLA for low-level action execution. It also builds the RoboPoly simulation benchmark to achieve long-term sequential collaboration operations of two robots under distributed local observation conditions.

Critical strengths
  1. The research topic is important. The issue of multi-robot cooperation has significant practical significance. In real scenarios, a large number of tasks rely on distributed multi-robots.
  2. The proposed method achieves a good balance between the high-level task planning reasoning and the robot's fine physical control capabilities through the collaborative work of the high-level VLM and the low-level VLA.
  3. The verification was conducted in two environments, RoboPoly and the modified RoboTwin. Through comprehensive ablation experiments, the effectiveness of the proposed method was proven.
Critical weaknesses
  1. Currently, it is only applicable to the dual-robot setup. The effectiveness of the framework under larger robot scales has not been verified yet, and its scalability needs further proof.
  2. What are the distinctive features of the proposed method compared to the hierarchical architecture model?
  3. All the experiments were conducted in simulations, lacking deployment verification on real robot hardware. The transferability from simulation to reality needs to be tested.
Questions for authors

See above Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Lf8R

We are grateful to Reviewer Lf8R for their constructive feedback and valuable suggestions. In this response, we address the reviewer's comments and provide further clarification on the key concerns raised.

Weakness 1

Currently, it is only applicable to the dual-robot setup. The effectiveness of the framework under larger robot scales has not been verified yet, and its scalability needs further proof.

Thank you for your constructive comment. DuoMind does not impose assumptions on the number of robots. Its distributed architecture has an orchestrator and an action model for each robot, and supports coordination through peer messages, making this framework applicable to any numbers of robots. Understandably, one key challenge lies in the significant compute resources required to demonstrate multi-agent coordination; as the the number of robots increases, the required compute resources (for demonstration generation, agent fine-tuning and evaluation) grow exponentially, which exceeds quickly what an academic lab could afford. Considering these requirements, we focus on two robots to initiate this line of research in this study, hoping to stimulate interest in this important area in upcoming multi-robot systems. We appreciate the reviewer’s suggestion and plan to extend RoboPoly with tasks involving more robots and evaluate DuoMind with larger robot sizes, in future work.

Weakness 2

What are the distinctive features of the proposed method compared to the hierarchical architecture model?

Thank you for the opportunity to clarify these distinctions. First, most existing hierarchical robot frameworks focus on single-agent settings. In a single-robot hierarchical architecture, System 2 reasons about the task and the consequences of the robot’s own behavior to guide low-level execution. In DuoMind’s distributed multi-robot setting, each orchestrator must additionally consider its peers’ behavior and its potential influence on the shared workspace before generating instructions to its own action model. To this end, our system design explicitly requires each orchestrator to reason over other agents' information and generate instructions that better harness its action model to achieve effective multi-robot coordination. Second, DuoMind integrates inter-agent communication into high-level orchestration. Each orchestrator reasons over incoming messages while generating both an instruction for its own action model and a compact natural-language message for its peers. This exchange enhances mutual awareness, mitigates action conflicts, and enables coordinated execution. Together, these nontrivial designs establish a promising paradigm for future research on distributed multi-robot systems.

Weakness 3

All the experiments were conducted in simulations, lacking deployment verification on real robot hardware. The transferability from simulation to reality needs to be tested.

Thank you for highlighting the importance of real-world validation. Real-world multi-robot experiments entail substantial resources. Collecting coordinated demonstrations through teleoperation by multiple operators is particularly demanding, and fine-tuning multiple agents increases computational requirements relative to single-robot experiments. Hardware evaluation also requires at least two robots and a carefully configured shared workspace. Given these practical constraints, the present study focuses on simulation to explore and get an in-depth understanding first. As noted above, our hope is to stimulate interest in this important area in upcoming multi-robot systems. It is nontrivial to develop the two-system framework to distributed multi-robot coordination, providing a foundation for subsequent real-world studies. We agree with the reviewer that physical validation is important, and as a rule of thumb, we carry out simulation studies to initiate this line of research.

Dear Reviewer Lf8R:

As the author–reviewer discussion period draws to a close today, we would appreciate any further thoughts you may have on whether our response addresses your main concerns.

As noted in our earlier response, our hope is to take initiatives to study multi-robot systems (which has barely been explored) and stimulate interest in this important area. It is nontrivial to develop the proposed two-system framework for distributed multi-robot coordination - it took us more than an year to work out this paper. We hope this study opens a door to this very important area in physical AI. Provided that you agree with this, we will appreciate you can improve your rating.

Reviewer_Lf8ROpenReview ↗

Thanks for your kind reply. I will keep my current review

Thanks for your kind reply. I will keep my current review

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all the reviewers for the constructive comments and effort in rebuttal.

We have revised our paper (highlighted in blue text color). The modifications are summarized as follows:

  1. We add ablation experiment using GPT-6 Astra as orchestrator in Appendix D. (In response to Reviewer Aadf and 9vFJ)
  2. We add ablation experiment that shows the effect of each component in shared message and DuoMind's robustness to realistic communication conditions in Appendix E. (In response to Reviewer Aadf and 9vFJ)
  3. We add ablation experiment using oracle orchestrator in Appendix F. (In response to Reviewer Aadf)
  4. We apologize for the misleading term "global camera," which we have now renamed "shared camera" throughout the paper.
AuthorsOpenReview ↗

General Response

As the author–reviewer discussion period draws to a close on Nov. 18, we would appreciate any further thoughts you may have on whether our response addresses your main concerns.

Thank you again for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.