acceptodds
Under review as a conference paper at ICLR 2027

MILD-MPC: Multi-Agent Imitation Learning via Distributed Model Predictive Control

Abstract

Controlling multi-agent systems in real time requires a scalable optimization method that satisfies constraints. However, existing distributed trajectory optimization methods typically assume fixed problem structures, and their solve times exceed the time budget as the number of agents grows. We propose Multi-agent Imitation Learning via Distributed Model Predictive Control (MILD-MPC), a framework for learning a permutation invariant predictive control policy that imitates a distributed MPC expert, generalizes to larger problems, and runs in a single forward pass at deployment. In addition, we develop Model Predictive Distributed Differential Dynamic Programming (MPD-DDP), a new distributed MPC method that handles time-varying problem structures, provides high-quality expert demonstrations, and serves as a backup policy when a safety verifier invalidates the learned policy's trajectory. We demonstrate that the deployed policy achieves substantial timing improvements over both a centralized MPC baseline and the MPD-DDP expert, generalizes to problems with more agents, and remains robust to agent failures during deployment, across 2D and 3D obstacle-field navigation tasks. In hardware experiments, a policy trained with 8 agents transfers to 15 robots without retraining, completing the task collision-free in real time.

Primary area:
applications to robotics, autonomy, planning
Keywords:
imitation learning, distributed optimization, multi-agent systems, optimal control, nonlinear control, model predictive control, robotics

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

This paper proposes MILD-MPC, a multi-agent imitation learning framework based on distributed model predictive control. The framework learns a policy from demonstrations provided by a distributed MPC expert, which also serves as a backup policy during deployment. By combining distributed MPC with imitation learning, it reduces online solve times and uses a safety verifier to assess the validity of predicted trajectories, providing an approach to accelerating real-time multi-agent control. The framework’s effectiveness is evaluated through comparative simulations and hardware experiments.

Critical strengths
  1. The proposed method offers an effective approach to accelerating real-time multi-agent control. Comparisons with the evaluated baselines demonstrate substantial reductions in online computation time, highlighting its practical potential.

  2. The paper evaluates the proposed framework through diverse simulation scenarios and real-robot experiments, providing evidence of its computational efficiency, transferability to larger teams, and robustness under the tested agent-failure scenarios.

Critical weaknesses
  1. In Fig. 2, the proposed MPD-DDP and learning augmented methods are compared against a centralized SQP baseline. However, distributed methods are generally intended to improve computational efficiency over centralized approaches, so this comparison does not sufficiently establish the proposed framework’s computational advantages over other representative distributed methods. The evaluation should include comparisons with existing distributed approaches, such as ADMM-DDP, which is discussed repeatedly in the paper.

  2. In the workflow shown in Fig. 1, trajectory validation is applied to the candidate trajectories generated by the learned policy and refined by the safety filter. However, no further validation is shown for the output of the MPD-DDP backup policy.

Questions for authors
  1. Fig. 2 shows that MPD-DDP has a higher average solve time with 12 agents than with 15 agents. What accounts for this non-monotonic trend? The paper does not provide an explanation.

  2. The maximum numbers of ADMM iterations for MPD-DDP are set to 5, 30, 20, 45, and 60 for the 8, 12, 15, 20, and 30 agent settings, respectively. How were these limits selected? In particular, why is the iteration limit higher for 12 agents than for 15 agents? The rationale for varying these limits is not sufficiently explained.

  3. Could the average solve time axis in Fig. 2 use a logarithmic scale? The smaller values are difficult to distinguish on the current scale, which limits the readability of the comparison.


Update: After reading the responses, I decided to keep the original score.

5 comments
AuthorsOpenReview ↗

Author response (1/n)

We thank the reviewer for acknowledging the practical potential of the proposed framework. We hope that our revised manuscript and our responses address the reviewer's questions and concerns.

Summary

We have clarified the following points.

  1. The comparisons with existing distributed approaches, such as ADMM-DDP.
  2. Why there is no further validation is shown for the output of the MPD-DDP backup policy.
  3. What accounts for the non-monotonic trend in the average solve time and the maximum number of ADMM iterations in Fig. 2 with respect to the number of agents? How these limits were selected.
  4. A logarithmic scale in Fig. 2 for better readability.

Our Response

1. (Weakness 1): In Fig. 2, the proposed MPD-DDP and learning augmented methods are compared against a centralized SQP baseline. However, distributed methods are generally intended to improve computational efficiency over centralized approaches, so this comparison does not sufficiently establish the proposed framework’s computational advantages over other representative distributed methods. The evaluation should include comparisons with existing distributed approaches, such as ADMM-DDP, which is discussed repeatedly in the paper.

Response:

We appreciate the reviewer's point that the comparison with existing distributed approaches, e.g., ADMM-DDP, should be included. Indeed, we agree that distributed methods generally improve the computational cost over centralized methods. In this context, we have compared our method with ADMM-DDP (labelled as Trajectory optimization) in Appendix I: ``What is the effect of dynamic neighborhood set reconstruction in time-varying problem structure?", where ADMM-DDP fails to complete the task safely in 8 and 12-agent settings due to the lack of dynamic neighborhood reconstruction (ADMM-DDP assumes fixed neighborhood across time). Furthermore, we wish to point that we have included the comparison of MILD-MPC with our proposed distributed MPC baseline, MPD-DDP, in Fig. 2.

AuthorsOpenReview ↗

Author response (2/2)

2. (Weakness 2): Trajectory validation is applied to the candidate trajectories generated by the learned policy and refined by the safety filter. However, no further validation is shown for the output of the MPD-DDP backup policy.

Response:

Indeed, it is possible that MPD-DDP returns an infeasible solution due to the time constraint or when the optimization problem in the intermediate step is infeasible. Checking the feasibility of the final trajectory is straightforward, and it boils down to the solution of the final backup policy or specification of the emergency behavior, e.g., emergency stop (e-stop). In our context, we chose to use MPD-DDP as a final backup policy, which handles constraints by construction. In terms of emergency behavior specification, we tested the quadrotor software-in-the-loop (SITL) experiment with e-stop enabled and the Robotarium experiment under the platform-wide e-stop guard. We have not included these engineering techniques in our methodology as they are out of scope for this work.

3. (Question 1): Fig. 2 shows that MPD-DDP has a higher average solve time with 12 agents than with 15 agents. What accounts for this non-monotonic trend? The paper does not provide an explanation. The maximum numbers of ADMM iterations for MPD-DDP are set to 5, 30, 20, 45, and 60 for the 8, 12, 15, 20, and 30 agent settings, respectively. How were these limits selected? In particular, why is the iteration limit higher for 12 agents than for 15 agents? The rationale for varying these limits is not sufficiently explained.

Response:

Thanks for sharing the astute observations! We wish to make a point that the non-monotonic trends of solve times and ADMM iteration count with respect to the number of agents are due to the nonconvexity of the problem. Fig. 2 task involves two moving obstacles. When these time-varying per-agent state constraints (gi(x~i)≤0,∀i∈[M]\mathbf{g}_i(\tilde{\mathbf{x}}_i) \le 0, \forall i \in [M]) are combined with the inter-agent state constraints (hia(x~ia)≤0,∀i∈[M]\mathbf{h}_i^a(\tilde{\mathbf{x}}_i^a) \le 0, \forall i \in [M]) in a receding horizon scheme, then the gap between the DDP trajectories (xi,ui\mathbf{x}_i, \mathbf{u}_i) and the safe trajectories (x~i,u~i\tilde{\mathbf{x}}_i, \tilde{\mathbf{u}}_i) each distributed optimal control problem at each replanning step is not necessarily monotonic with respect to MM. Since ADMM enforces the consensus between the DDP trajectories and the safe trajectories, the non-monotonic gap between these trajectories implies that the ADMM iteration count is not necessarily monotonic, i.e., a larger gap needs more ADMM iterations to close. Since the solve time of MPD-DDP mostly depends on the ADMM iteration count, the non-monotonic ADMM iteration count with respect to MM implies the non-monotonicity of the MPD-DDP solve time.

The median values of the primal residual (∣x~i−xi∣F\\|\tilde{\mathbf{x}}_i - \mathbf{x}_i\\|_F) at ADMM iteration 1 across all replanning steps for M=(8,12,15,20,30)M = (8, 12, 15, 20, 30) in Fig. 2 are (0.006,0.030,0.021,0.031,0.057)(0.006, 0.030, 0.021, 0.031, 0.057), which show the non-monotonic gap between the DDP trajectories and the safe trajectories with respect to MM. This result also aligns with the trends of the maximum number of ADMM iterations (5,30,20,45,60)(5, 30, 20, 45, 60) and the solve times (0.026,0.185,0.178,0.746,1.27)(0.026, 0.185, 0.178, 0.746, 1.27)s of MPD-DDP with respect to M=(8,12,15,20,30)M = (8, 12, 15, 20, 30).

4. (Question 2): Could the average solve time axis in Fig. 2 use a logarithmic scale? The smaller values are difficult to distinguish on the current scale, which limits the readability of the comparison.

Response:

Thanks for pointing out the readability of Fig. 2! We have added another version of Fig. 2 in which the y-axis is in a log scale.

AuthorsOpenReview ↗

Author response

We appreciate the reviewer for the constructive feedback. We updated our manuscript to address the reviewer's points.

Reviewer_f569OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a multi-agent distributed optimal control framework. They present a new solver Model Predictive Distributed Differential Dynamic Programming (MPD-DDP), and its novelty is that it supports time-varying neighborhoods. They use MPD-DDP, during offline process, to generate data for imitation learning, and, during inference, as a backup policy when downstream safety checks fail. They propose safety filters and verifiers for the imitation learning. They validate on quadrotor simulation experiments and on robotorium hardware experiments.

Critical strengths

The evaluation of the method is thorough. The trade-off between latency, speedup and verifier acceptance in Fig. 2 is interesting and validates the proposed architecutre.

Critical weaknesses
  • The area of distributed optimal control is mature and well understood, which reduces the novelty of the first contribution: “We introduce Model Predictive Distributed DDP (MPD-DDP), a scalable distributed nonlinear MPC method that supports time-varying neighborhoods.” Any model-predictive control method that replans at runtime can support neighborhoods that change between solver calls, including references already in the related work [Liu et al. 2024]. The paper positions MPD-DDP against Liu et al. (2024) with the statement “yet the subgraph structure is updated only between episodes rather than within the horizon,” but this positioning is not convincing. In Sec. 2.2 and Algorithm 1 (Appendix E), the authors state that the neighborhood sets “are recomputed at each replanning step,” and the neighborhood stays fixed within each solve. Can the authors please clarify what “within the horizon” means, and the difference between a replanning step and a timestep? Usually MPC planners replan at every timestep, so these are the same in my understanding. If MPD-DDP only recomputes the graph between solver calls, it is not clear what it adds beyond running ADMM-DDP in a receding-horizon loop. This is important because MPD-DDP is the main methodological (not architectural) contribution of the paper, and its novelty is not clearly stated

  • The papers is spread thin over multiple areas, preventing going deep on a core contribution of the paper; the authors present results in: (i) numerical optimization (ii) transformer architecture and (iii) policy architecture. This makes it unclear what the deep conceptual insight or novelty we are supposed to understand from the method. It also makes it hard to develop any one idea sufficiently, e.g. the policy architecture idea of using imitation learning then safety filter with backup is known and well studied (e.g. [Riviere2020]).

Minor:

  • the method appears overengineered, e.g. why do both the safety filter and the verification gate effectively control the conditions: b < 0 and g < 0?
  • This spread results in a lot of notation that makes the paper hard to read.
Questions for authors

Please let me know if I misunderstood any critical weaknesses.

6 comments
AuthorsOpenReview ↗

Author response (1/n)

We thank the reviewer for acknowledging that the evaluation of the proposed method is thorough. We hope that our revised manuscript and our responses address the reviewer's questions and concerns.

Summary

We have clarified the following points.

  1. Positioning of the proposed distributed MPC method, MPD-DDP, against Liu et al. 2024 [1].
  2. Novelty of MPD-DDP.
  3. The difference between a replanning step and a time step.
  4. The deep conceptual insight or novelty of the proposed method.
  5. Why both the safety filter and the verification gate effectively control the conditions: b<0b < 0 and g<0g < 0.
AuthorsOpenReview ↗

Author response (2/n)

Our Response

1. (Weakness 1): Any model-predictive control method that replans at runtime can support neighborhoods that change between solver calls, including references already in the related work [1]. The paper positions MPD-DDP against Liu et al. (2024) with the statement “yet the subgraph structure is updated only between episodes rather than within the horizon,” but this positioning is not convincing.

Response:

We thank the reviewer for pointing out that the positioning against [1] is not convincing. Indeed, our wording “yet the subgraph structure is updated only between episodes rather than within the horizon.” is misleading, and thus we have removed the statement. Nevertheless, the proposed distributed MPC method (MPD-DDP) is still different from [1] in the neighborhood computation scheme. In Algorithm 3, Graph Evolution for CAVs in [1], the entire graph is partitioned into disjoint subgraphs, where each subgraph corresponds to the neighborhood set. This algorithm is inherently centralized, whereas the neighborhood reconstruction in MPD-DDP is fully decentralized. One of the major advantages of fully decentralized neighborhood reconstruction approach is that it opens the possibility to cover partially observable settings, e.g., each agent forms their own neighborhood sets based on their observation even when the central subgraph allocator is not available.

2. (Weakness 1): If MPD-DDP only recomputes the graph between solver calls, it is not clear what it adds beyond running ADMM-DDP in a receding-horizon loop. This is important because MPD-DDP is the main methodological (not architectural) contribution of the paper, and its novelty is not clearly stated.

Response:

We thank the reviewer for sharing their keen insight. We agree that the area of distributed optimal control is mature. We wish to make the point that MPD-DDP aims to build a global controller from multiple local controllers [2], and receding horizon control is a mechanism to realize this goal. One of the main motivations for using multiple local controllers in this work is to relax the fixed neighborhood assumption. The reasoning is that time-varying neighborhoods emerge as multiple local controllers, each with different sets of fixed neighborhoods, are synthesized into a global controller.

However, this problem is not trivial, as the global value function for each agent cannot be approximated by the existing initialization strategies that employ time-shifted policies of the local controllers [2, 3] if the problem structures are not fixed across the local controllers in distributed optimal control. Ideally, the local value function models of the local controllers in standard DDP [4] do not model the same global function, since the value function depends on time kk, denoted by a sequence (V(x,k))k=1T(V(\mathbf{x}, k))_{k=1}^T, where TT is a fixed finite time horizon. Meanwhile, the time-dependent local value function V(x,k)V(\mathbf{x}, k) collapses to a global, time-independent value function V(x)V(\mathbf{x}) as T→∞T \to \infty [5, Section 1.8].

To approximate the time-independent global value function in a computationally efficient way, Tassa et al. [2] propose a fixed-length non-discounted receding horizon scheme that constructs local controllers at each time instant, initialized with the policy of the local controller from the previous time step. However, this approach becomes ineffective if the problem structure (neighborhood set) varies across local controllers, since the value function of MPD-DDP depends on the neighborhood set Ni\mathcal{N}_i and the augmented safe variable (Appendix D.2), based on Problem 2.

Suppose that the value function for the centralized formulation in Problem 1, denoted as V(x,k)V(\mathbf{x}, k), converges to the time-independent value function V(x)V(\mathbf{x}) as T→∞T \to \infty with a non-empty set of admissible inputs. Suppose that the global value function of the centralized problem can be decomposed as V(x)=∑i∈[M]Vi(xi)V(\mathbf{x}) = \sum\limits_{i \in [M]} V_i(\mathbf{x}_i), where Vi(xi)V_i(\mathbf{x}_i) is the global value function of agent ii. To approximate Vi(xi)V_i(\mathbf{x}_i) for each ii, MPD-DDP initializes the local controllers with the time-shifted trajectories of primal, dual, and global variables in ADMM and zeroes out dual variables when the neighborhood set changes at each replanning step.

Therefore, we wish to make a point that synthesizing global controllers for multiple agents from local controllers in distributed optimal control is not trivial, and the novelty of MPD-DDP lies in the global value function approximation approach, which initializes ADMM with time-shifted primal, dual, and global variables of local controllers, and resets the dual variables depending on the change in the neighborhood set.

AuthorsOpenReview ↗

Author response (3/3)

3. (Weakness 1): What is the difference between a replanning step and a timestep? Usually MPC planners replan at every timestep, so these are the same in my understanding.

Response:

In our framework, we wish to point out that we consider mm-step MPC [5], where m≥1m \ge 1 is called the control implementation horizon, i.e., we reoptimize every mm time step.

4. (Weakness 2): The papers is spread thin over multiple areas, preventing going deep on a core contribution of the paper; the authors present results in: (i) numerical optimization (ii) transformer architecture and (iii) policy architecture. This makes it unclear what the deep conceptual insight or novelty we are supposed to understand from the method. It also makes it hard to develop any one idea sufficiently, e.g. the policy architecture idea of using imitation learning then safety filter with backup is known and well studied (e.g. [Riviere2020]).

Response:

We thank the reviewer for sharing constructive feedback. We wish to suggest that the main novelty lies in our systematic approach to bridging two distinct areas: Multi-agent imitation learning and distributed MPC. In particular, our conceptual insights are as follows: i) existing approaches for approximating the global value function of the global controller using time-shifted policies of local controllers is not directly applicable if the problem structure is not fixed across local controllers in distributed optimal control, ii) the time-shifted trajectories of dual variables in distributed optimization in particular are not valid for initialization when the problem structure changes, iii) MPD-DDP approximates the global value function for multiple agents by initializing ADMM using time-shifted trajectories of primal, dual, global variables of local controllers and resetting dual variables when the neighborhood set changes, and iv) bridging distributed optimal control with multi-agent imitation learning leads to an effective framework for real-time multi-agent control. To our knowledge, MILD-MPC is the first framework that bridges multi-agent imitation learning with distributed nonlinear MPC, backed by empirical evidence for end-to-end real-world deployability via hardware tests.

5. (Minor 1): Why both the safety filter and the verification gate effectively control the conditions: b<0b < 0 and g<0g < 0.

Response:

We wish to make a point that the formulation of the safety filter problem in Problem 3 imposes per-agent control and state constraints bi(u~i)≤0\mathbf{b}_{i} (\tilde{\mathbf{u}}_{i}) \le 0, gi(x~i)≤0\mathbf{g}_{i} (\tilde{\mathbf{x}}_{i}) \le 0, for all agents i∈[M]i \in [M]. If Problem 3 is feasible, the solution to Problem 3 satisfies the constraints: bi(u~i)≤0\mathbf{b}_{i}(\tilde{\mathbf{u}}_{i}) \le 0, gi(x~i)≤0\mathbf{g}_{i}(\tilde{\mathbf{x}}_{i}) \le 0. The verification gate additionally checks the constraint violations of the safety filter solution in case of infeasibility.

Reviewer_7zxoOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses multi-agent navigation and collision avoidance in continuous spaces under nonlinear dynamics. The authors propose MILD-MPC, a framework that trains a shared, permutation-invariant Transformer to imitate a distributed MPC expert, MPD-DDP, which extends ADMM-DDP with neighborhood reconstruction at each replanning step. At deployment, the learned policy predicts control sequences that are refined by a predictive safety filter and checked by a verifier; the distributed optimizer remains part of the deployed system and serves as a backup when verification fails. Experiments on unicycle and quadrotor systems report computational speedups and collision-free task completion, while hardware experiments demonstrate deployment of a policy trained with 8 agents on 15 physical robots. I consider this physical deployment a meaningful strength, as it demonstrates that the integrated framework can operate beyond simulation. However, the experiments do not sufficiently isolate how much of the acceleration comes from learning, conditional optimization, or changes to controller settings. Although the paper provides architecture and optimizer comparisons, task-completion evidence, and hardware timing statistics, it does not establish a systematic quality–computation advantage under matched settings or against a strong compatible external controller. The contribution of the individual learning and safety components to the deployed system also remains insufficiently characterized. Overall, the paper demonstrates a functioning learning-augmented controller with meaningful hardware validation, but the evidence does not yet convincingly establish the benefit of the proposed combination over well-tuned alternatives.

Critical strengths

The paper provides meaningful end-to-end deployment evidence.

The Robotarium experiment demonstrates that the integrated controller can operate on physical robots, including an eight-to-fifteen-agent transfer without retraining the neural policy. The paper reports collision-free completion over three repetitions and provides solve-time and verifier-acceptance statistics. The quadrotor simulation also includes state estimation, sensor emulation, and an agent-failure scenario. These experiments are consequential because they go beyond simple prediction accuracy and demonstrate that the proposed components can be integrated into an operating control system.

Critical weaknesses
  1. The evaluation does not sufficiently isolate the reported acceleration or establish competitiveness under matched conditions.

    • The central comparison changes the optimization problem as well as the controller. Appendix I.2 uses Ri=500diag(1,0.1)R_i=500 \mathrm{diag}(1,0.1) for standalone MPD-DDP but Ri=100diag(1,0.1)R_i=100 \mathrm{diag}(1,0.1) for the learning-augmented controller in Figure 2. The quadrotor comparisons also change state and terminal costs, obstacle margins, and inter-agent safety distances. The controllers additionally use different ADMM iteration limits, which the authors attribute to the benefit of learned warm-starting. These choices may be appropriate for independently tuned systems, but they do not cleanly isolate the contribution of learning. The missing control is a comparison of conventionally warm-started MPD-DDP, learned warm-starting with an MPC solve at every replanning step, and the complete pipeline under the same objectives and constraints. Different iteration counts are legitimate, provided that comparable solution quality and constraint satisfaction are demonstrated.
    • The external comparisons are too narrow to establish the broader quality-computation advantage. The principal comparisons are against centralized SQP and the authors’ own expert. At least one strong external controller should be evaluated on a representative setting with compatible dynamics, obstacle handling, information access, and online replanning requirements. The literature in [1–4] provides potential planning and control comparisons.
    • The paper does report safe task completion, illustrative completion times, and hardware solve-time statistics. What remains missing is a systematic comparison of common closed-loop quality measures - such as trajectory cost, completion time (flowtime and makespan), and minimum separation - alongside computation time under matched conditions. A breakdown of the costs of neural prediction, filtering, verification, and fallback would clarify the source of the acceleration. Because hardware execution is asynchronous, the authors should also explain how execution proceeds during long solves and how these delays affect task performance.
    • The transfer experiment does not separate policy transfer from controller retuning. The network trained with eight agents is successfully deployed with larger teams, but the safety-filter horizon, verifier threshold, and some obstacle margins vary with team size. At 30 agents, verifier acceptance is 50.5%. The authors should explain how the team-dependent settings were selected and report task quality and fallback computation across team sizes. This would distinguish generalization of the learned policy from the contribution of optimization and target-setting tuning.
  2. The contribution of the individual components to the deployed system remains insufficiently established.

    • The architecture experiment evaluates the learned policies in isolation, not their benefit within the deployed hybrid. Table 2 provides useful evidence of robustness to token reversal and compares unshielded policy performance. However, it removes the predictive safety filter and backup, so it does not establish the architecture’s effect on the complete system’s quality–computation trade-off. The consequential comparison is the proposed Transformer against an existing simpler baseline, such as the sum-pooling Deep Sets model, using the same training data, comparable training and tuning budgets, and the same filter, verifier, and backup. Reporting task quality, end-to-end runtime, and fallback frequency would reveal whether the stronger raw policy translates into a better deployed controller. The zero-completion results for Deep Sets and the RNN also warrant clearer training and tuning details. Targeted ablations of the state-matching and constraint-loss terms, and of the predictive filter while retaining the verifier and backup, would further clarify the individual contributions.
    • Safety attribution and fallback failure handling remain unclear. The predictive filter deliberately omits inter-agent constraints, leaving their checking to the verifier and subsequent handling to the backup optimizer. Appendix I.9 additionally reports a platform-level control barrier function for the supplementary hardware experiments, but does not quantify its interventions or clearly establish whether it was active in the main 15-robot experiment. The authors should specify where this layer was active and report its intervention frequency and correction magnitude. Separately, Algorithm 1 applies controls after a bounded ADMM loop without showing an explicit final feasibility check. The behavior when the backup returns an insufficiently feasible solution, or when a solve is delayed, should be explained. These omissions limit attribution of the observed collision-free performance to the proposed mechanisms.
    • Proposition 1 does not supply the missing justification. Its proof establishes permutation properties and support for variable token counts, then invokes Cybenko’s theorem without showing why it applies to the implemented architecture. Those properties alone do not establish universal approximation. The proposition needs a precise approximation domain, cardinality assumptions, and an argument connecting this architecture to an applicable theorem. Even a valid universal-approximation result would not establish that this architecture is preferable to simpler policies or that it generalizes effectively to larger teams; those remain empirical questions.

References

[1] Shaoul, Yorai, et al. “Multi-Robot Motion Planning with Diffusion Models.” ICLR. 2025.

[2] Dergachev, Stepan, and Konstantin Yakovlev. “Decentralized Uncertainty-Aware Multi-Agent Collision Avoidance With Model Predictive Path Integral.” IROS. 2025.

[3] Zhang, Yuhao, et al. “D4orm: Multi-Robot Trajectories with Dynamics-aware Diffusion Denoised Deformations.” IROS. 2025.

[4] Dergachev, Stepan, et al. “CoRL-MPPI: Enhancing MPPI With Learnable Behaviours For Efficient And Provably-Safe Multi-Robot Collision Avoidance.” arXiv:2511.09331. 2025.

Questions for authors
  1. Matched experimental controls. Under identical objectives and constraints, how do conventionally warm-started MPD-DDP, learned warm-starting with an MPC solve at every replanning step, and the complete pipeline compare? Please if available report collision-free task completion, common trajectory-quality measures, and end-to-end computation time. This comparison would distinguish gains from learned initialization and conditional execution from those attributable to controller retuning.

  2. Competitive and architectural comparisons. Can you provide a representative comparison against a strong compatible external controller, and a full-pipeline comparison between the proposed Transformer and the existing simpler Deep Sets / RNN policy?

  3. Training, deployment, and transfer protocol. Please specify the numbers of independent training, validation, and test problems and training seeds, what was randomized or held out, and whether separate policies were trained for each dynamics model or scenario family. For transfer to larger teams, how were the team-dependent filter horizons, verifier thresholds, and safety margins selected - using separate validation scenarios or the reported test scenarios?

  4. Please also clarify whether the future obstacle positions supplied in Appendix F are known or predicted at deployment.

17 comments
AuthorsOpenReview ↗

Author response (1/n)

We thank the reviewer for emphasizing the strength of the proposed MILD-MPC framework as ``The paper provides meaningful end-to-end deployment evidence" and for acknowledging our hardware experiments and quadrotor software-in-the-loop (SITL) experiment with state estimation, sensor emulation, and an agent-failure scenario. We hope that our revised manuscript and our responses address the reviewer's questions and concerns.

Summary

We added the following experiments (item 1-6) to address the reviewer's points and evaluated the results using the closed-loop metrics (trajectory cost, flow time, makespan, and minimum separation) and analyzed the solve time breakdown. In addition, we added our clarifications regarding the following points (item 7-11).

  1. A comparison of conventionally warm-started MPD-DDP, learned warm-starting with an MPC solve at every replanning step, and the complete pipeline under the same objectives and constraints.
  2. A comparison of our proposed methods with a strong external controller, MPPI-ORCA, in an online replanning setting.
  3. A full pipeline comparison between neural network architectures.
  4. An ablation study of state-matching loss, constraint loss, and the safety filter, while retaining the verifier and backup policy.
  5. An empirical analysis of our hardware experiments regarding how execution proceeds during long solves, how these delays affect task performance, and the platform-level control barrier function's intervention frequency and correction magnitude.
  6. A performance analysis of the proposed methods in an additional experiment on unknown future obstacle positions.
  7. How the team-dependent filter horizons, verifier thresholds, and safety margins were selected - using separate validation scenarios or the reported test scenarios.
  8. Training and tuning details for zero-completion results (Deep sets, RNN).
  9. Why there is no feasibility check for the backup policy in the final algorithm and the behavior when the backup trajectory is insufficiently feasible, or the solve is delayed.
  10. Why Proposition 1 applies to the implemented neural network architecture.
  11. The numbers of independent training, validation, and test problems, and training seeds, what was randomized or held out, and whether separate policies were trained for each dynamics model or scenario family.
AuthorsOpenReview ↗

Author response (2/n)

Our response

1. (Weakness 1-1, Weakness 1-3, Question 1): A comparison of conventionally warm-started MPD-DDP, learned warm-starting with an MPC solve at every replanning step, and the complete pipeline under the same objectives and constraints. A systematic comparison of common closed-loop quality measures - such as trajectory cost, completion time (flowtime and makespan), and minimum separation - alongside computation time under matched conditions. A breakdown of the costs of neural prediction, filtering, verification, and fallback would clarify the source of the acceleration.

Response: Thanks for the good point.

Table A

Mmethodcost ↓flow time (s) ↓makespan (s) ↓closest obstacle (m) ↑closest inter-agent (m) ↑TC ↑Safe TC ↑
8MPPI-ORCA211940 ± 1254018.76 ± 1.3924.29 ± 2.190.004\color{red}{0.004}0.35038/4031/40\color{red}{31/40}
MPD-DDP (conventional WS)12165712.9813.460.1700.3168/88/8
MPD-DDP (learned WS)12886013.9215.910.2590.2068/88/8
MILD-MPC (no safety filter)13123614.6716.760.2330.2538/88/8
MILD-MPC (full pipeline)12866613.9616.200.2600.2658/88/8
12MPPI-ORCA258399 ± 1830619.02 ± 1.0926.40 ± 0.000.003\color{red}{0.003}0.35345/6034/60\color{red}{34/60}
MPD-DDP (conventional WS)13827612.4022.180.2800.26212/1212/12
MPD-DDP (learned WS)13230111.5215.020.2780.21312/1212/12
MILD-MPC (no safety filter)12628011.9315.350.154\color{red}{0.154}0.12812/1211/12\color{red}{11/12}
MILD-MPC (full pipeline)13283411.5716.140.1810.19612/1212/12
15MPPI-ORCA320401 ± 245517.05 ± 0.0726.40 ± 0.000.014\color{red}{0.014}0.35157/7548/75\color{red}{48/75}
MPD-DDP (conventional WS)16141911.7214.650.1840.18915/1515/15
MPD-DDP (learned WS)15583711.1413.560.2540.27415/1515/15
MILD-MPC (no safety filter)15463111.7715.250.1630.25015/1515/15
MILD-MPC (full pipeline)15288611.1411.980.1600.26115/1515/15
20MPPI-ORCA888316 ± 8970920.99 ± 0.3926.40 ± 0.000.022\color{red}{0.022}0.35154/10050/100\color{red}{50/100}
MPD-DDP (conventional WS)36039015.1120.690.1750.16820/2020/20
MPD-DDP (learned WS)32264513.2815.020.1940.19620/2020/20
MILD-MPC (no safety filter)31995114.0618.150.152\color{red}{0.152}0.23620/2019/20\color{red}{19/20}
MILD-MPC (full pipeline)31253513.2415.110.1690.20420/2020/20
30MPPI-ORCA2505325 ± 14996523.74 ± 0.5026.40 ± 0.000.002\color{red}{0.002}0.35069/15065/150\color{red}{65/150}
MPD-DDP (conventional WS)99188718.2021.550.1830.17830/3030/30
MPD-DDP (learned WS)95611917.7822.670.2370.15830/3030/30
MILD-MPC (no safety filter)96190517.5619.570.1800.106\color{red}{0.106}30/3028/30\color{red}{28/30}
MILD-MPC (full pipeline)94458717.4822.410.1650.23430/3030/30
AuthorsOpenReview ↗

Author response (3/n)

Table B

Mmethodtotal (s) ↓average (ms) ↓NN prediction ↓safety filter ↓verification ↓ADMM ↓other ↓acceptance rate (%) ↑
8MPPI-ORCA3.734.7 ± 0.0——————
MPD-DDP (conventional WS)3.5918.0 ± 1.9———17.11.0—
MPD-DDP (learned WS)8.2741.6 ± 5.16.6—4.029.71.2—
MILD-MPC (no safety filter)1.608.0 ± 0.14.7—2.50.00.9100.0
MILD-MPC (full pipeline)1.959.8 ± 0.14.74.00.20.01.0100.0
12MPPI-ORCA3.734.7 ± 0.0——————
MPD-DDP (conventional WS)32.44163.0 ± 10.9———161.81.2—
MPD-DDP (learned WS)16.5683.2 ± 4.57.4—4.470.21.2—
MILD-MPC (no safety filter)5.6028.1 ± 1.36.6—3.916.61.181.5
MILD-MPC (full pipeline)8.1741.1 ± 3.36.811.30.221.51.364.0
15MPPI-ORCA3.764.7 ± 0.0——————
MPD-DDP (conventional WS)34.23172.0 ± 4.6———170.41.6—
MPD-DDP (learned WS)15.3176.9 ± 5.711.6—7.656.21.6—
MILD-MPC (no safety filter)6.0530.4 ± 1.37.9—4.716.61.384.0
MILD-MPC (full pipeline)7.3737.0 ± 2.18.111.90.215.41.480.5
20MPPI-ORCA4.085.1 ± 0.0——————
MPD-DDP (conventional WS)137.74692.1 ± 14.5———690.41.8—
MPD-DDP (learned WS)28.66144.0 ± 15.412.3—7.9122.01.8—
MILD-MPC (no safety filter)15.4777.7 ± 9.99.4—5.960.91.557.0
MILD-MPC (full pipeline)17.2786.8 ± 2.010.345.20.329.11.858.0
30MPPI-ORCA4.105.1 ± 0.0——————
MPD-DDP (conventional WS)245.561234.0 ± 52.1———1231.82.2—
MPD-DDP (learned WS)37.64189.1 ± 3.812.0—8.5166.42.2—
MILD-MPC (no safety filter)26.59133.6 ± 9.911.5—8.0112.02.239.5
MILD-MPC (full pipeline)23.59118.5 ± 1.811.423.30.581.02.354.0

In Table A and Table B, we report closed-loop task performance and breakdown of solve time, comparing the conventional and learned warmstart strategies (WS) in MPD-DDP under the same objectives and constraints. In Table A, MPD-DDP (learned, conventional WS) consistently achieves full safe task completion across all MM. In Table B, learned WS shows approximately 2.31×2.31 \times slower average solve time (41.641.6 ms) compared to the conventional WS (18.018.0 ms) on M=8M=8, but shows approximately (1.96,2.24,4.81,6.53)×(1.96, 2.24, 4.81, 6.53) \times faster average solve times (83.283.2, 76.976.9, 144144, 189.1189.1) ms compared to the conventional WS (163.0,172.0,692.1,1234.0163.0, 172.0, 692.1, 1234.0) ms on M=(12,15,20,30)M=(12, 15, 20, 30). Specifically, the learned WS accelerates ADMM computation times by roughly (2.30,3.03,5.66,7.40)×(2.30, 3.03, 5.66, 7.40)\times faster than the conventional WS on M=(12,15,20,30)M=(12, 15, 20, 30).

Cost function and constraint parameters. All methods use the same cost function and the constraint parameters in the same problem MM. Cost functions for all MM settings and for all methods are selected as follows: Qi=15⋅diag⁡([1,1,0])Q_i = 15 \cdot \operatorname{diag}([1, 1, 0]), Ri=500⋅diag⁡([1,0.1])R_i = 500 \cdot \operatorname{diag}([1, 0.1]), Qif=10000⋅diag⁡([1,1,0.1])Q_i^f = 10000 \cdot \operatorname{diag}([1, 1, 0.1]). Constraint parameters for each MM are selected as follows. The obstacle and the inter-agent safety distances are i) do=0.2d_o = 0.2, dsafe=0.35d_{\text{safe}} = 0.35 for M∈{8,12,15,30}M \in \{8, 12, 15, 30\}, and ii) do=0.26d_o = 0.26, dsafe=0.36d_{\text{safe}} = 0.36 for M=20M =20.

Comments on the paper

1 comment

We appreciate the reviewers for their meaningful feedback, including the thorough evaluation of the method (all reviewers) and its practical potential to accelerate real-time multi-agent control (Mwkw). We believe our proposed framework, MILD-MPC, dramatically improves the scalability of real-time multi-agent control under challenging state and control constraints by bridging multi-agent imitation learning and distributed model predictive control.

Main concerns are about the breadth of the scope (f569), novelty of the distributed MPC method (f569), comparison with other decentralized methods (Mwkw, 7zxo), and systematic empirical analysis (7zxo).

In our revision, we have clarified all questions, providing empirical analysis and clarifications to the concerns.


Summary

1. Novelty of the proposed method

We have updated the methodological novelty of the proposed distributed MPC method (MPD-DDP) in Section 2.2 (Model Predictive Distributed Differential Dynamic Programming), highlighting the initialization approach in distributed optimal control when the optimization problem potentially changes over time.

2. Empirical analysis and comparison with other decentralized methods

We have conducted additional experiments to analyze the individual modules in the proposed framework and added analysis in Appendix I.10-16. Specifically, we have shown the following results.

  1. The learned warmstart achieves meaningful speedup over the conventional warmstart in MPD-DDP (Appendix I.11).
  2. The proposed MILD-MPC consistently shows safe task completion results across settings with various numbers of agents, while the external decentralized baseline results in unsafe outcomes (Appendix I.13).
  3. The implemented Permutation invariant Transformer model achieves complete safe task completion results across settings with a range of agents, while other neural network architectures (RNN, Deep Sets) fall short in safe task completion results.
  4. Only the full loss function and the full pipeline with the safety filter show complete safe task completion results (Appendix I.14).
  5. The safety filter improves the safe task completion results in the full pipeline (Appendix I.12)
  6. Empirical computation time analysis in Tables 7 and 9 (Appendix I.10, 11).

3. More clarifications

In addition, we have clarified the following points.

  1. The task performance of the proposed methods when the future obstacle positions are unknown in Appendix I.15.
  2. Data generation details in Appendix I.5.
  3. Analysis of the non-monotonic trends of ADMM iteration count and average solve times with respect to the number of agents in Appendix I.8.

During the rebuttal process, we have tried to address all the questions and points raised in the reviews. We appreciate the reviewers for devoting the time to provide feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.