- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces SynerMind, a framework for long-horizon collaboration among heterogeneous robots. Its main contributions are a multimodal model for planning, execution, and reflection, a closed-loop workflow that coordinates robot-specific skills and supports failure recovery, and a collaboration-oriented training strategy. The framework is validated through decision evaluations, ablation studies, and simulated and real-world multi-robot experiments, showing improved task success and robustness over existing baselines.
Critical strengths
1.Clear motivation and coherent presentation: The paper is generally well structured and easy to follow. The motivation is convincing.
2.Methodological innovation: The proposed approach is technically sound and demonstrates a certain degree of methodological innovation.
Critical weaknesses
1.Incomplete specification of the method: several central components of the proposed method are not specified, limiting reproducibility in the claimed sources of performance improvement. For example, Eq. (2) does not provide the value or selection criterion for the weighting coefficient , and essential training configurations for this stage are missing.
2.The descriptions in the main text and the appendix are not fully consistent: In Eq. (1), the Planner takes only the instruction I and current state s_t as inputs, whereas Algorithm 1 additionally includes the current workflow and corrective feedback h. This discrepancy makes the Planner interface ambiguous and reduces the formal rigor and reproducibility of the proposed method.
Questions for authors
1.Clarify how you determine the value and search range of lambda in Eq. (2). If it is selected empirically, please add a sensitivity analysis in the experimental section and explain the selection criterion.
2.There is an inconsistency between Eq. (1) in the main text and Line 4 of Algorithm 1 in the appendix.
3.The construction and update mechanism of the state buffer B in Eq. (8) is not specified.
4.The Planner-Execution-Reflection architecture closely resembles existing agentic workflows, and its architectural novelty is not sufficiently established.
3 comments
Q4 Technical Contribution
Q4 Technical Contribution
We thank the reviewer for the meaningful suggestion. The main technical contributions are as follows.
1. Causal Logic driven Collaborative Chain-of-Thought.
Existing agentic frameworks primarily prompt general-purpose models (ChatGPT-4o) to to generate task assignments and retries execution based on failure feedback. In contrast, we organize task execution constraints into a collaborative chain of thought (CoT). Training establishes a structured mapping from task states to robot-skill decisions. Each CoT trace identifies unfinished subtasks from verified execution history and checks their execution preconditions. It then matches executable subtasks to robot capabilities and derives the corresponding skill invocations, while identifying the causes of failure and revising execution plans accordingly. We construct SynerAgentic-102k from real robot execution trajectories to provide task-state supervision for this reasoning process as shown below, while Appendix I demonstrates dynamic plan adjustment based on identified failure causes, in contrast to static retries.
| Model | Response |
|---|---|
| COHERENT | <think> Bottle is delivered. Rock is at (2.40, -1.10). The mobile robot is at (0.80, 0.30). Collect the rock next. </think> |
| SynerMind-Brain (after training) | <Global Task>Complete the object manipulation task by sorting the delivered bottle and collecting the rock.</Global Task> <Current Observation>The bottle has been successfully delivered and is ready for sorting. The bottle has not yet been sorted, and the rock has not yet been collected. The next planned actions are to sort the bottle and approach the rock.</Current Observation> <Robot States>Fixed Dual-Arm Robot: Target object is bottle, target region_xy is [0.35, 0.20], execution status is pending. Mobile Manipulation Robot: Navigation target pose is [2.30, -0.95, 1.57], execution status is pending.</Robot States> <Verified History>{"bottle_delivered":true,"bottle_sorted":false,"rock_collected":false}</Verified History><Reflection>Errors: omitted bottle sorting; unmet navigation precondition for rock collection. Assign bottle sorting; schedule navigation before collection.</Reflection> |
2. A Multi-Agent System Based on Collaborative-State Mining.
Existing agentic frameworks primarily rely on predefined prompts to construct static workflows. In contrast, rather than manually specifying how Planner, Reflection, and Execution should collaborate, we leverage agentic RL training to enable our policy to learn how to coordinate decision-making across different stages through interactions with actual tasks. Specifically, we propose CSM-GRPO, a reinforcement learning algorithm driven by collaborative state mining. The trained shared policy enables dynamic coordination among the planning, reflection, and execution agents, thereby improving the overall task execution performance of multiple robots on long-horizon tasks.
To evaluate the independent contribution of SynerMind-Flow, we compare different collaborative workflows on four representative real-world tasks, varying only the collaborative workflow.
| Method | T2 | T8 | T14 | T20 | Avg. SR (%) |
|---|---|---|---|---|---|
| PEFA-style Flow | 6/10 | 5/10 | 5/10 | 6/10 | 55.0 |
| SynerMind-Flow | 9/10 | 9/10 | 8/10 | 9/10 | 87.5 |
Thank You for Your Consideration
Dear Reviewer We3w,
Thank you again for the valuable and constructive feedback.
As the author-reviewer discussion will close soon, we kindly ask if our response have addressed your concerns. Please feel free to let us know if there are any other question remaining.
Best,
Paper 8646 Authors
W1 & Q1 Cold Start Training Details, W2 & Q2 Planner Interface Clarification and Q3 Buffer Clarification.
We thank the reviewer for the meaningful suggestions and address each concern below.
W1 & Q1 Cold Start Training Details
We further clarify the rationale for selecting the loss weight λ in Eq. (2) and the training configuration for the cold-start stage.
1. Joint Optimization and Loss Balancing. In Eq. (2), ℒaware provides fundamental state awareness, while ℒagent focuses on more complex collaborative reasoning and skill decisions. We therefore favor assigning a higher weight to the latter. Based on the Validation Risk Minimization theory[1,2], we had conducted a sensitivity analysis over the range [0.5,2.0]. By jointly evaluating Awareness Acc., R-L, and Agentic Acc., we select λ=1.5 to balance state awareness and collaborative decision-making performance.The table below presents our sensitivity analysis, which we will include into Section 5.3.4 of the revised manuscript.
2. Cold-Start Training Configuration. In Stage 1, we use Qwen3.5-9B as the base model and perform joint supervised fine-tuning with LoRA on NVIDIA A100 80GB GPU. The detailed configuration is provided below, and we will include the rationale for the loss weight selection and the detailed training configuration into Appendix B.1 of the revised manuscript.
References
[1] Hu, Y., Xian, R., Wu, Q., Fan, Q., Yin, L., & Zhao, H. (2023). Revisiting Scalarization in Multi-Task Learning: A Theoretical Perspective. NeurIPS 2023.
[2] Li, Y., Liu, Z., & Xing, E. (2025). Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models. ICML 2025.
W2 & Q2. Planner Interface Clarification. Eq. (1) lists two main inputs to the Planner, the task instruction I and the current collaborative state st. Line 4 of Algorithm 1 in the appendix specifies the full set of inputs, which also includes the existing workflow Π and corrective feedback h. In the revised manuscript, we will explicitly list all four inputs in Eq. (1) to align it with Algorithm 1.
Q3 Buffer Clarification
We provide details on buffer initialization, online difficulty updates, and the training sampling procedure. We will include these details into Appendix B.2.1 of the revised manuscript.
1. Buffer Initialization. We construct the state buffer (B) from the training split of SynerAgentic-102k. Each entry represents a collaborative decision state containing multimodal observations, robot states, task instructions, workflow progress, and verified execution history, providing the collaborative context for reward computation.
2. Online Difficulty Update. During CSM-GRPO training, each sampled state generates K=8 candidate responses. We compute their rewards using Eq. (5) and normalize them to [0,1] as defined in Appendix C. The state difficulty δ(s) is then computed using Eqs. (6) and (7). Whenever a state is resampled and a new round of response evaluation is completed, its difficulty score in the buffer is updated based on the current policy's responses. Unsampled states retain their most recent difficulty estimates.
3. Difficulty-Aware Sampling. In Eq. (8), we mix the uniform sampling distribution pB(s) over the buffer with a state difficulty-based Softmax sampling distribution, using ρ=0.5 and τtemp=1.0 for state sampling in subsequent training batches.