acceptodds
Under review as a conference paper at ICLR 2027

Geometric Pareto Control: Physics-Supervised Pareto Representation Learning via Riemannian Energy-Gradient Flow

Abstract

We study multi-objective sequential control problems in physical systems whose dynamics and operational constraints are known or can be represented by accurate physics-based models. Although reinforcement learning has been widely explored in such settings, standard policy learning still faces high-dimensional action search, fixed or externally supplied objective trade-offs, retraining under changed priorities, and hard feasibility requirements at deployment. Our premise is that known physics and constraints expose a compressible action structure that can be learned as a reusable response map before deployment. The resulting Pareto-relevant response family is organized before deployment, so online control begins with a physics-supervised prediction and local correction instead of repeatedly solving a nonlinear program. We propose Geometric Pareto Control (GPC), which embeds the supported family of dynamically feasible Pareto-optimal control responses into a continuous latent homotopy space using offline scalarized optimization under the known dynamics. At deployment, the measured physical state supplies priorities to the learned map; guarded local updates and domain-specific refinement adapt its decoded action or plan. We state formal assumptions under which decoded actions remain feasible and the action error does not accumulate over the rollout horizon. Across analytical control, safe multi-agent navigation, and optimal power flow, GPC compares favorably with optimization and safe-RL baselines in safety, feasibility, and real-time multi-objective performance. These results support physics-supervised learning of reusable Pareto-relevant responses that can be adapted online under changing priorities.

Primary area:
learning on time series and dynamical systems
Keywords:
multi-objective control, Pareto representation learning, physics-supervised learning, Riemannian optimization, geometric control

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_54oFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a new framework for multi-objective control under known physical dynamics and constraints denoted Geometric Pareto Control. The idea is to amortize constrained optimization though generation of scalarized Pareto-relevant control solutions offline and use them to learn a latent representation. Such a latent representation is then decoded and produces the control actions. During deployment, urgency weights (which are dependent on the state) specify the current objectives, while local correction and feasibility checks are used to adapt the responses.

Dual-encoder representation learning, rollout-aware supervision, and domain-specific online refinement are part of the proposed methodology. For the optimal power flow setting, the authors also introduce a Riemannian energy-gradient flow with RK2 integration and a Lie-local correction. Furthermore, they establish conditional feasibility and non-cumulative matched-action error bounds. The proposed framework shows favorable online efficiency, constraint satisfaction, and objective performance when evaluated on analytical control, multi-agent navigation, and AC optimal power flow experiment.

Critical strengths

1. Strong empirical performance and practical motivation.

The reported results are promising across different control domains. In particular, GPC achieves objective values comparable to MPC in the IEEE 30-bus OPF experiments while substantially reducing online computation time. The navigation experiments also shows strong safety and task-completion performance. Overall these results provide meaningful evidence for the practical value of amortizing constrained control optimization with a similar framework.

2. Integration of physical constraints and an explicit theoretical analysis.

The proposed framework combines offline constrained optimization, learned response representations, and online physical correction. The authors also provide a detailed analysis of feasibility and matched-action error, including the effects of localization, decoder sensitivity, and final output transformations.

Critical weaknesses

1. Limited comparison with closely related approaches and fairness for baselines evaluation

GPC is closely related to amortized constrained optimization, as it relies on offline numerical solutions to supervise a learned control-response map. However, the experimental evaluation lacks a closely matched supervised amortized-optimization baseline, such as a preference-conditioned neural solution map with equivalent training information, online refinement, and feasibility correction. This makes it difficult to determine whether the reported improvements arise from the proposed framework or from the more general benefit of amortizing expensive optimization procedures. Moreover, since the method already incorporates differentiable physical rollouts during its second training stage, a comparison with differentiable predictive control (DPC) would be informative. Such a comparison could help clarify the motivation for relying on optimizer-generated supervision rather than directly optimizing control objectives through differentiable dynamics.

The interpretation of the existing comparisons is confounded by differences in computational budgets and training information. Whole the authors address (at least partially) candidate-count fairness in navigation, GPC receives perturbation-aware training and network parameter descriptors that are not provided equivalently to the RL baselines in the perturbed OPF experiment. These differences limit the conclusions that can be drawn about comparative robustness and the advantages of the proposed learning framework.

2. Insufficient experimental isolation of the proposed geometric components

Even if the overall effectiveness of GPC is well established by the reported results, the numerical experiments do not convincingly define how much of the performance improvement can be attributed to the learned Pareto geometry, Riemannian metric, and Lie-local correction, as opposed to supervised optimization and subsequent online refinement.

The authors provide an auxiliary RK2-versus-RK2+Lie comparison in Appendix J.4, but the highlighted improvement in objective value seems to be modest, since both variants achieve full recorded feasibility in the auxiliary experiments. Moreover the navigation experiments rely on learned plan correction and finite candidate refinement rather than the explicit RK2/Lie update, making it hard to attribute their strong performance to the proposed Riemannian navigation mechanism.

More systematic component-level ablations, particularly comparisons between Euclidean and metric-preconditioned updates, with and without Lie-local corrections, and with and without geometric representation regularization, would make more clear the contribution of every single component. As currently presented, the results demonstrate the effectiveness of the complete pipeline more convincingly than the incremental value of its geometric mechanisms.

3. Limited connection between the theoretical geometric and feasibility claims and their empirical verification

While the paper introduces a continuous latent homotopy representation of Pareto-relevant control responses, the mathematical properties of this representation are not sufficiently characterized. In particular, the proposed representation losses encourage local continuity without establishing that the learned representation preserves the topology of the constrained solution family or provides continuous transitions between distinct Pareto-optimal responses. This distinction is especially important in nonconvex control problems, where feasible solution families may contain disconnected branches.

Similarly, the main feasibility theorem relies on uniform Lipschitz bounds, positive constraint margins, and bounded localization and update errors. Although the theoretical argument is valid under the stated assumptions, it remains unclear to what extent these conditions are quantitatively verified for the evaluated controllers. This is particularly relevant when optimal solutions lie on active constraint boundaries, where strict feasibility margins may vanish, or when the deployed controller incorporates potentially discontinuous candidate-selection operations.

Consequently, while the theoretical analysis provides useful conditional guarantees, stronger evidence connecting its assumptions to the implemented controllers would be a great addition to properly establish their practical applicability and to really substantiate the claimed geometric interpretation.

Questions for authors
  1. Comparison with amortized optimization and differentiable predictive control. Given that GPC uses offline optimizer-generated supervision and differentiable physical rollouts, how would it compare against a preference-conditioned amortized optimizer or an end-to-end differentiable predictive controller? A comparison using matched model capacity, physical information, training resources, and final feasibility mechanisms would materially clarify the contribution of the geometric representation.
  2. Contribution of the geometric components. Could the authors provide a more systematic ablation of the latent geometry, Riemannian metric, and Lie-local correction? The auxiliary RK2/Lie results are useful, but it remains unclear whether these components provide substantial improvements over simpler latent gradient updates or non-geometric learned response correction.
  3. Fairness and computational accounting. Could the authors clarify the performance under matched end-to-end online computation budgets, beyond matched candidate counts? In addition, how sensitive are the OPF comparisons to the different training information available to GPC and the RL baselines? Reporting total offline reference-generation and training costs would also help quantify the practical amortization trade-off.
5 comments
AuthorsOpenReview ↗

1. How does GPC compare with a preference-conditioned amortized optimizer and differentiable predictive control?

Reviewer comment — Critical Weaknesses 1 (first part):

Limited comparison with closely related approaches and fairness for baselines evaluation....

Reviewer comment — Questions For Authors 1:

Comparison with amortized optimization and differentiable predictive control.....

Authors’ response.

With the physics, projection, safety templates, and 4,096-candidate refinement shared, all three reach 100% success in moving scenes, and GPC attains a lower mean cost than the supervised map when priorities change externally (3988.25 versus 4045.59).

New in the revision. To the best of our knowledge, no existing method performs amortized optimization for Pareto-optimal control with priorities that change online, so we built the two closest controls ourselves:

  • a supervised amortized map trained on the same optimizer data as GPC;
  • native-objective DPC trained through the differentiable horizon-8 rollout without action labels, with parameter count matched to GPC within 0.14%.

Neither includes GPC's dual encoder, measurement-based localization, or latent updates. Results for every scene and priority setting, with mean decision times over the complete decision path, are in Tables 10–11.

What GPC adds. All three amortize optimization and share the final mechanisms. GPC represents the response across priorities and links it to localization, accepted updates, and feasibility margins, so that the validity of the response can be checked as conditions change. The representation itself is tested separately (new in the revision):

SceneFull modelState-side representation removedChange
Static1.2391321.286480+3.8210%
Moving1.1305061.169993+3.4929%

Why optimizer supervision rather than DPC alone. DPC optimizes the objective directly through the rollout. Optimizer supervision gives GPC verified reference responses across priorities. In the submission, localization and the feasibility margins of Theorem 1 are defined with respect to these references, and Proposition 4 obtains the margins from verified anchor responses.

Evidence: Table 10 (p. 43); Table 11 (p. 43); Table 9 (p. 43); Proposition 4 (p. 18).

AuthorsOpenReview ↗

2. What do the latent geometry, Riemannian metric, and Lie-local correction each contribute?

Reviewer comment — Critical Weaknesses 2:

Insufficient experimental isolation of the proposed geometric components....

Reviewer comment — Questions For Authors 2:

Contribution of the geometric components. Could the authors provide a more systematic ablation of the latent geometry, Riemannian metric, and Lie-local correction? The auxiliary RK2/Lie results are useful, but it remains unclear whether these components provide substantial improvements over simpler latent gradient updates or non-geometric learned response correction.

Authors’ response.

We verified the contribution of each component by varying it alone, and the results below confirm that each one improves the controller; the dual representation gives the largest gain.

Component varied aloneResultSource
Dual representationStage-I objective 3.8210% (static) and 3.4929% (moving) worse without itnew, Table 9 (p. 43)
Riemannian metricConditional gap 0.269712% → 0.227032%new, Table 20 (p. 50)
Lie-local correctionConditional gap 0.227032% → 0.221944%new, Table 20 (p. 50)
Lie-local correctionPost-Newton AC mismatch below 10−510^{-5} → below 10−610^{-6}in the submission, Appendix J (p. 36)
Lie coefficient60/60 feasible steps at every value testednew, Table 21 (p. 51)
Rollout-aware stagesMoving scenes: 19/20 → 20/20 safe successes; cost 2521.99 → 2400.54new, Table 12 (p. 44)

The metric and Lie rows are measured in OPF, where the RK2/Lie flow runs online. All of GPC's OPF results in the paper were produced with this flow running online, including the Lie-local term: with it, the executed dispatch reaches the 10−610^{-6} tolerance of the IPOPT solver, whereas without it the maximum AC mismatch stays an order of magnitude higher (below 10−510^{-5}). Navigation decodes its plans from the learned response geometry, whose contribution is the dual-representation row. The comparison with and without locality regularization is in Appendix M.4.

Evidence: Table 9 (p. 43); Table 20 (p. 50); Appendix J (p. 36); Table 12 (p. 44).

AuthorsOpenReview ↗

3. How do online computation, OPF training information, and offline cost affect the comparisons?

Reviewer comment — Critical Weaknesses 1 (second part):

The interpretation of the existing comparisons is confounded by differences in ....

Reviewer comment — Questions For Authors 3:

Fairness and computational accounting. Could the authors clarify the performance under matched end-to-end online computation budgets, beyond matched candidate counts? In addition, how sensitive are the OPF comparisons to the different training information available to GPC and the RL baselines? Reporting total offline reference-generation and training costs would also help quantify the practical amortization trade-off.

Authors’ response.

They do not change the conclusions. Every method is now timed over its complete decision path, and the RL baselines receive GPC's per-step information. Offline cost is reported as the workload of every stage, with a break-even point of about 3,050 decisions per hour of offline computation.

  • Online computation (new in the revision). Navigation times include projection, safety templates, and final refinement (Table 10). In Case118, median end-to-end time including final correction is 35.5 ms for GPC versus 1,615.2 ms for NMO-OPF (IPOPT).

  • Training information (new in the revision). With GPC's per-step weights and an AC-PF correction, TD3 reaches reference gaps of 0.5046% / 10.4191% and CPO-style PPO-Lagrangian 5.5193% / 9.2208% (nominal / perturbed), versus 0.3011% / 0.2496% for GPC.

  • Descriptors (new in the revision). On 4,994 validation paths, the network descriptor carries structured information that GPC uses without requiring exact matching:

    Descriptor inputChange in Stage-I objective
    Held nominal+0.022015%
    Shuffled+0.027539%
    Removed (zero)+386.134358%
  • Offline cost (new in the revision). Navigation uses 60,000 static and 120,000 moving-scene labels, and the nominal OPF references cover 5000 loads × 50 weights. With the Case30 decision times in Table 2 (1193.8 ms for LRH-MPC, 12.3 ms for GPC), each hour of offline computation is recovered after about 3,050 decisions.

Evidence: Table 10 (p. 43); Table 17 (p. 48); Table 14 (p. 46); Table 15 (p. 46); Table 22 (p. 51); Appendix M.13 (p. 52).

Reviewer_ZLgHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns a reusable control-response map from offline scalarized constrained optimization. At deployment, state-dependent priorities guide response adaptation. OPF uses RK2/Lie latent updates, while navigation uses learned plan correction and finite candidate refinement. Experiments cover analytical control, multi-agent navigation, and AC OPF. Conditional feasibility and matched-reference action-error bounds are also provided.

Critical strengths

The navigation performance is substantial and survives a candidate-count control: with 384 candidates for both methods, GPC achieves 96/100 safe successes versus 35/100 for NMPC (Appendix I.4). This provides convincing evidence for the usefulness of the learned planning pipeline. The OPF results demonstrate a useful quality–latency tradeoff, achieving costs and feasibility close to MPC with roughly 97× lower decision latency (Table 2). The additional 100-sequence evaluation supports consistent objective quality.

Critical weaknesses

The strongest experiments do not validate the mechanism foregrounded by the paper. The title and Sections 3.4–3.5 emphasize Riemannian energy-gradient flow and RK2/Lie updates, whereas navigation uses learned plan correction and structured candidate refinement without executing that flow online (Appendix F.5). This distinction is disclosed, but it leaves a gap between the central methodological story and its empirical support. The matched-candidate result validates the learned planning pipeline; it does not establish the contribution of the geometric mechanism. A supervised plan initializer with identical correction and refinement is the crucial missing comparison.

The main guarantee provides limited method-specific theoretical insight. Its feasibility conclusion follows from Lipschitz continuity, bounded localization/update errors, and an assumed uniform positive margin of the final decoded responses. This argument applies to any bounded update and does not essentially use Pareto structure or the geometric integrator. The difficult step, establishing these margins and error bounds for the deployed controller, is not demonstrated. Likewise, the horizon-independent action bound follows from uniform per-cycle assumptions and compares against a matched reference, rather than controlling accumulated physical trajectory error.

Questions for authors

How does GPC compare with a standard supervised plan initializer using the same reference data and identical online correction/refinement, under matched runtime budgets? Beyond the auxiliary study in J.4, which results establish a benefit from the Riemannian/RK2/Lie mechanism over an equally expensive Euclidean correction? Can the deployment guarantee be instantiated with verified margins and localization/error bounds, yielding a nonzero certified update radius on an evaluated region?

3 comments
AuthorsOpenReview ↗

1. How does GPC compare with a supervised plan initializer under identical correction and refinement?

Reviewer comment — Critical Weaknesses 1:

The strongest experiments do not validate the mechanism foregrounded by the paper....

Reviewer comment — Questions For Authors 1:

How does GPC compare with a standard supervised plan initializer using the same reference data and identical online correction/refinement, under matched runtime budgets?

Authors’ response.

Under identical correction, projection, safety templates, and 4,096-candidate refinement, with the complete decision path timed, both reach 100% success in moving scenes, and GPC attains the lower mean cost when priorities change externally (3988.25 versus 4045.59 for the supervised initializer).

New in the revision. To the best of our knowledge, no existing method performs amortized optimization for Pareto-optimal control with priorities that change online, so we built the closest amortized baseline ourselves. It is a supervised plan initializer trained on the same reference data and observable inputs as GPC and deployed with GPC's final mechanisms. It does not include GPC's dual encoder, measurement-based localization, or latent updates.

Scene, prioritySuccess, GPC / supervised (%)Mean cost, GPC / supervisedMean time, GPC / supervised (ms)
Static, fixed90 / 952175.28 / 2146.36118.07 / 117.83
Static, external90 / 1004932.58 / 4858.54118.38 / 118.32
Moving, fixed100 / 1002402.11 / 2294.74121.55 / 121.28
Moving, external100 / 1003988.25 / 4045.59120.11 / 119.05

What GPC adds. Both are amortized, and this comparison measures closed-loop quality under identical final mechanisms. The supervised initializer returns one plan per input. GPC represents the response across priorities and links it to localization, accepted updates, and feasibility margins, so that the validity of the response can be checked as conditions change. The representation itself is tested separately (new in the revision): the Stage-I objective is 3.8210% / 3.4929% worse without the state-side representation. In the submission, Appendix I.4 reported that at an equal candidate count (384), GPC attains 96/100 safe successes versus 35/100 for multistart NMPC.

Where the geometric mechanism is tested. The geometric mechanism has two parts, and each is tested where it runs.

  • The learned response geometry is used in every domain; navigation decodes its plans from it, followed by learned plan correction and finite refinement (Appendix F.5 (p. 30) of the submission). The representation test above (3.8210% / 3.4929%) isolates its contribution.
  • The RK2/Lie flow runs online in OPF. All of GPC's OPF results in the paper were produced with this flow running online, including the Lie-local term: with it, the executed dispatch reaches the 10−610^{-6} tolerance of the IPOPT solver, whereas without it the maximum AC mismatch stays an order of magnitude higher (below 10−510^{-5}). In Case30, GPC's reference gap (0.3011% / 0.2496%) is close to that of NMO-OPF (IPOPT) (0.2752% / 0.2281%), at 12.060 / 12.600 ms per decision instead of 139.130 / 139.931 ms; on 46 held-out Case118 topologies, it attains 99.24% online feasibility. New in the revision, with only the latent update changed, the conditional reference gap falls from 0.269712% (Euclidean RK2) to 0.227032% (Riemannian RK2) and 0.221944% (RK2+Lie) (Table 20 (p. 50)).

Evidence: Table 10 (p. 43); Table 9 (p. 43); Appendix I.4 (p. 35); Table 20 (p. 50); Table 14 (p. 46); Table 16 (p. 47); Appendix J (p. 36).

AuthorsOpenReview ↗

2. What does the Riemannian/RK2/Lie update add over an equally expensive Euclidean correction?

Reviewer comment — Questions For Authors 2:

Beyond the auxiliary study in J.4, which results establish a benefit from the Riemannian/RK2/Lie mechanism over an equally expensive Euclidean correction?

Authors’ response.

In OPF, the Riemannian metric and the Lie term each lower the reference gap relative to the Euclidean RK2 update, and the Lie term lowers the AC mismatch by an order of magnitude.

New in the revision. With only the latent update changed on the perturbed network:

Latent updateConditional reference gap (%)
Direct decoding0.370414
Euclidean RK20.269712
Riemannian RK20.227032
RK2+Lie0.221944

Table 20 reports the per-step CPU time of each update, and all four Lie coefficients tested give 60/60 feasible steps.

In the submission. Appendix J reports that the Lie term lowers the maximum post-Newton AC mismatch from below 10−510^{-5} to below 10−610^{-6}; the revision adds 30 load trajectories to this check. This gain is significant: the AC mismatch is power that the physical network cannot balance, so OPF solvers drive it to numerical precision, and the IPOPT solver in our OPF experiments uses a tolerance of 10−610^{-6}. The Lie term brings the executed dispatch to this tolerance.

Evidence: Table 20 (p. 50); Appendix J (p. 36); Table 21 (p. 51).

AuthorsOpenReview ↗

3. Can the guarantee be instantiated with a nonzero certified radius, and what is specific to GPC?

Reviewer comment — Critical Weaknesses 2:

The main guarantee provides limited method-specific theoretical insight. Its feasibility ...

Reviewer comment — Questions For Authors 3:

Can the deployment guarantee be instantiated with verified margins and localization/error bounds, yielding a nonzero certified update radius on an evaluated region?

Authors’ response.

Yes. The revision gives a numerical certificate for the convex analytical controller: on the reported closed-loop rollouts, the common radius exceeds the localization distance at all 20 cycles.

In the submission. Theorem 1 gives a nonzero radius r⋆=min⁡iγi/(LiK)r^\star=\min_i\gamma_i/(L_iK) whenever the margins γi\gamma_i are positive, and Proposition 4 obtains these margins from verified anchor responses.

New in the revision. Appendix M.14 checks the condition numerically, using the paper's reference fiber (the complete saved reference collection) and certified constants, including K≤12.7003K\le12.7003 for the final map with clipping. For example, r⋆≥1.3096r^\star\ge1.3096, while the localization distance is at most 0.33120.3312.

What is specific to GPC. The condition is stated for GPC's final executed map, including correction and candidate selection. Proposition 1 bounds the RK2 and Lie-local displacement inside the radius. Because the controller re-localizes from the measured state every cycle, the action-error bound does not grow with the horizon (Eq. 16).

Further closed-loop evidence. The other evaluated closed loops also keep positive margins and feasible final responses:

  • In the submission, Table 5 reports navigation minimum margins averaged per episode: scene 0.301 (static) and 0.203 (moving), and agent–agent 2.477 and 1.065.
  • In the submission, the Case30 final response is feasible in 300/300 and 299/300 steps.
  • New in the revision, the final response is feasible in 99.24% of 2,760 steps on 46 held-out Case118 graphs, without retraining.

Evidence: Theorem 1 (p. 6); Proposition 4 (p. 18); Appendix M.14 (p. 52); Table 24 (p. 53); Proposition 1 (p. 6); Table 5 (p. 35); Table 16 (p. 47).

Reviewer_64owOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces Geometric Pareto Control (GPC) for fast multi-objective control in systems with known physics and constraints. Offline constrained optimization generates a family of control responses, from which GPC learns a compact representation. Online, it adapts the decoded action or plan using priorities derived from the measured physical state.

The stated contributions are: (i) dual-encoder representation learning with neighborhood regularization and rollout-aware refinement; (ii) online adaptation through guarded RK2/Lie latent updates for OPF and learned plan correction with finite candidate refinement for navigation; (iii) conditional feasibility and matched-action error guarantees; and (iv) reuse across unseen parameter realizations within the modeled family without retraining.

Experiments cover analytical control with convex and nonconvex objectives, multi-agent navigation in static and moving scenes, and AC optimal power flow under nominal conditions, parameter perturbations, and a held-out topology change. The reported results indicate a favorable balance among objective value, feasibility, and latency, although GPC is not best on every metric. Additional evidence includes repeated analytical trials, GPC-only evaluations over 100 load sequences per OPF setting, a navigation comparison with equal candidate counts, and a limited OPF study isolating the Lie correction.

Critical strengths
  1. Practical problem and coherent framework. The paper addresses fast multi-objective control under changing priorities and coupled physical constraints. Combining offline constrained response generation, a learned compact representation, and online correction offers a practical way to reuse expensive optimization results across operating conditions.
  2. Broad evaluation with useful controls. Experiments span analytical control, multi-agent navigation, and AC optimal power flow, covering distinct objectives, constraints, and response representations, including single actions and multi-step plans. Repeated analytical trials, GPC evaluations over 100 load sequences in each of the nominal and branch-perturbed OPF settings, and a moving-scene navigation comparison with equal candidate counts provide evidence beyond individual illustrative trajectories. Parameter perturbations and a held-out topology change further test reuse beyond nominal operating conditions.
  3. Explicit and carefully scoped theory. The deployment certificate states its conditions explicitly, accounting for localization error, total accepted displacement, decoder sensitivity, and feasibility margins. The manuscript distinguishes empirical feasibility from formal certification and discusses limitations involving equality residuals, changing contexts, numerical references, and the gap between decoded-response feasibility and physical safety.
  4. Attention to deployment details. The framework incorporates measurement-based re-localization, update guards, and application-specific correction or refinement as parts of the deployed controller. It also distinguishes the RK2/Lie update used for OPF from the learned plan correction and candidate refinement used for navigation.
Critical weaknesses
  1. Engagement with prior work. The paper insufficiently engages with closely related literature. Section 2 cites no work published after 2023. Although Appendix B includes some recent references, neither discussion addresses [i]–[viii], which overlap with important aspects of GPC, including parameter-conditioned Pareto learning, adaptive and safety-augmented approximate MPC, geometric feasibility correction, and constrained manifold learning. The authors should clarify GPC’s methodological differences from these approaches and substantiate its advantages through representative benchmarks where applicable.
  2. Novelty: The manuscript appears to build its contributions on established ideas, including offline Pareto control construction (e.g., [vii]), parameter-conditioned Pareto learning (e.g., [i, ii]), adaptation without retraining (e.g., [iii]), and neural or geometric methods with online feasibility mechanisms (e.g., [iv, v]). These precedents do not invalidate GPC’s specific contributions, but the manuscript does not sufficiently distinguish its proposed constructions from the closest alternatives. In particular, the methodological advances offered by the dual-encoder training, urgency-driven correction, and deployment certificate require clearer explanation.
  3. MPC Latency: The unusually high MPC latencies warrant clarification. Although MPCs can be computationally expensive, runtime also depends strongly on solver choice, initialization, derivative computation, stopping criteria, and implementation. The manuscript mentions IPOPT, but the implementation details of the MPC baselines remain unclear. Please provide these details, together with the hardware specifications and timing scope. A comparison with an established MPC implementation using sequential quadratic programming (SQP), warm starting, a real time iteration (RTI) scheme, and an efficient QP solver would strengthen the computational claims.
  4. Limited evidence of Lie-local correction: Appendix J.4 provides limited evidence for the added value of the Lie correction. It changes the RK2 latent code by a mean ℓ2 distance of 3.56×10−53.56\times10^{-5}, while the reported mean cost decreases from 282.57 to 282.00 (≈0.2%) with unchanged recorded feasibility. The reported maximum AC mismatch bounds improve from below 10−510^{-5} to below 10−610^{-6}, but it is unclear whether these residuals are measured before or after any subsequent power-flow correction, so the practical significance of this improvement for the executed dispatch is unclear. Quantitative cost comparisons are provided for only one of the five evaluated maps, on a single trajectory, using checkpoints distinct from the main OPF benchmark. Paired results across all maps and multiple load trajectories, together with the additional computation time per step, would better establish this component’s practical contribution.

References

[i] J. Cheng et al., “Parametric Pareto Set Learning: Amortizing Multi-Objective Optimization with Parameters,” IEEE TEVC, 2025.

[ii] J. Cheng et al., “Parametric Pareto Set Learning for Expensive Multi-Objective Optimization,” AAAI, 2026.

[iii] H. Hose et al., “Parameter-adaptive approximate MPC: Tuning neural-network controllers without retraining,” L4DC, 2024.

[iv] H. Hose et al., “Approximate Nonlinear Model Predictive Control With Safety-Augmented Neural Networks,” IEEE TCST, 2025.

[v] E. Liang et al., “Homeomorphic Projection to Ensure Neural-Network Solution Feasibility for Constrained Optimization,” JMLR, 2024.

[vi] L. Chen et al., “Efficient First-Order Optimization on the Pareto Set for Multi-Objective Learning under Preference Guidance,” ICML, 2025.

[vii] S. Ober-Blöbaum et al., “Explicit multiobjective model predictive control for nonlinear systems with symmetries,” International Journal of Robust and Nonlinear Control, 2021.

[viii] S. Gupta et al., “Learning a neural Pareto manifold extractor with constraints,” UAI, 2022.

Questions for authors
  1. Could the authors benchmark GPC against [i] and [vii] on suitable tasks, and include targeted comparisons with [iii], [iv], or [v] where applicable? Please compare solution quality, constraint satisfaction, and total deployment latency, including correction and refinement, under matched information and computational budgets.
  2. I could not find any code with the submission. An anonymized repository containing the evaluated implementations, configurations, and reproduction instructions would improve confidence in the comparisons.
  3. Theorem 1 applies Theorem C.1 at every cycle with common constants, but does not establish that its assumptions propagate from an initially feasible state. A nonempty reference fiber, positive margins, and bounded localization error must remain valid as the context changes, including the previous executed action for ramp constraints. Appendix E.1 acknowledges that the reference fiber may be empty. The contextual cover in Appendix C.3 applies when every current pair (c,z)(c,z) lies within (qz,qc)(q_z,q_c) of a verified pair, but does not establish that closed-loop evolution preserves this condition. Could the authors provide sufficient conditions for such propagation or explicitly distinguish the result from recursive feasibility? The distinction between feasibility of the evaluated response and actual plant safety, acknowledged in Appendix C.4, should also be stated alongside Theorem 1.
  4. In the construction preceding Eq. (13), each rotation axis vbv_b is parallel to the corresponding block of dd whenever the eigenvalues within that block coincide, making that block’s Lie-local correction vanish. Could the authors clarify the geometric rationale for this construction and assess whether approximate alignment contributes to the small latent-code difference reported in Appendix J.4?
8 comments
AuthorsOpenReview ↗

1. How does GPC differ from references [i]–[viii]?

Reviewer comment — Critical Weaknesses 1:

Engagement with prior work. The paper insufficiently engages with closely related literature....

Reviewer comment — Critical Weaknesses 2:

Novelty: The manuscript appears to build its contributions on established ideas...

Authors’ response.

We have cited [i]–[viii] and compared GPC with the closest of them. New in the revision, Section 2 cites [i]–[viii], Appendix M.15 discusses each, and within the rebuttal period we compared GPC directly with the most recent closest methods we could find: PPSL [i], explicit MOMPC [vii], and HoP [v] on the analytical tasks, and, for [iii, iv], a supervised solution map with GPC's final mechanisms and acados SQP/SQP-RTI.

Contribution. GPC's methodology differs from [i]–[viii] in one construction, defined in Sections 3–4 of the submission: it links the response representation, measurement-based localization, accepted updates, and feasibility margins, so that the validity of a learned response can be checked as conditions change. The three elements the review names serve this construction:

  • Dual-encoder training aligns the offline, physics-supervised response representation with a representation computed from observations, so the controller localizes itself from measurements at every cycle. Removing the state-side representation worsens the Stage-I objective by 3.8210% (static) and 3.4929% (moving).
  • Urgency-driven correction sets the objective priorities from the measured physical state and moves the response along the learned family as they change (Figure 3); every priority mode attains 3,000/3,000 feasible steps.
  • The deployment certificate is stated for the final executed map, including correction and candidate selection, and gives a checkable radius; the revision verifies it numerically for the convex analytical controller (r⋆≥1.3096r^\star\ge1.3096 versus a localization distance of at most 0.33120.3312).

In this respect GPC follows a common pattern in which the contribution lies in how established components are combined for a specific problem, as in AlphaGo (Silver et al., Nature 2016), DQN (Mnih et al., Nature 2015), and Diffusion Policy (Chi et al., RSS 2023).

Relation to each reference.

  • [i, ii] Parametric Pareto learning learns solution families over parameters and preferences. GPC adds localization and accepted updates for sequential physical control. On the analytical tasks, PPSL fails on 51% / 29% of trajectories versus 2% for GPC.
  • [iii, iv] Approximate MPC adapts a learned controller through parameter sensitivities [iii] or a safety augmentation [iv]. GPC organizes responses across objective priorities and evaluates a final executed map.
  • [v, vi, viii] Feasibility and Pareto geometry construct feasibility-preserving or Pareto-constrained maps. GPC couples such a representation to repeated localization from the physical state.
  • [vii] Explicit MOMPC precomputes a Pareto-action library. GPC replaces the library with a learned representation; explicit MOMPC fails on 41% / 33% of trajectories versus 2%.

Evidence: Section 2 (p. 2); Appendix M.15 (p. 53); Table 3 (p. 10).

AuthorsOpenReview ↗

Additional Evidence

Merged into the response above.

AuthorsOpenReview ↗

2. How does GPC compare with the closest methods on representative benchmarks?

Reviewer comment — Questions For Authors 1:

Could the authors benchmark GPC against [i] and [vii] on suitable tasks, and include targeted comparisons with [iii], [iv], or [v] where applicable? Please compare solution quality, constraint satisfaction, and total deployment latency, including correction and refinement, under matched information and computational budgets.

Authors’ response.

Yes. As requested, we have benchmarked GPC against [i] and [vii] and included targeted comparisons for [iii], [iv], and [v]. The results are summarized below.

New in the revision. PPSL [i], explicit MOMPC [vii], and HoP [v] were adapted to the analytical tasks of Table 3, with 100 trajectories per objective and the same safety and slew checks for every method. A trajectory counts as feasible only if it completes all 20 steps and passes these checks.

MethodReference gap, convex / nonconvex (%)Failure, convex / nonconvex (%)Time, convex / nonconvex (ms)
GPC0.043 / 0.0502 / 20.78 / 0.97
PPSL [i]2.24 / 1.28 (one-step)51 / 290.32 / 0.35
Explicit MOMPC [vii]1.77 / 1.51 (one-step)41 / 330.04 / 0.04
  • Solution quality. GPC's reference gap is 0.043% / 0.050%, versus 1.28–2.24% for PPSL and explicit MOMPC under their one-step reference.
  • Constraint satisfaction. GPC fails on 2% of trajectories, versus 29–51% for PPSL and explicit MOMPC; HoP [v] completes 93/100 and 94/100 trajectories.
  • Deployment latency. Every method decides in under 1 ms per step on these tasks.
  • [iii, iv]. The closest controls are a supervised solution map deployed with GPC's correction and refinement, timed over the complete decision path (Tables 10–11), and acados SQP/SQP-RTI (Table 13). In closed-loop navigation, GPC reaches 90–100% safe joint success versus 0–20% for acados.

Evidence: Table 3 (p. 10); Table 23 (p. 51); Table 10 (p. 43); Table 13 (p. 45).

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

The paper proposes Geometric Pareto Control (GPC) for multi-objective control of physical systems with known dynamics and constraints. Offline, it solves many constrained scalarized problems and learns an encoder–decoder map of feasible, Pareto-relevant responses; online, it sets objective priorities from physical urgency, applies a guarded local latent correction, and decodes a feasible action. Experiments on analytical control, multi-agent navigation, and AC optimal power flow show cost close to MPC at roughly 100× lower latency.

Critical strengths
  1. The problem is practically relevant, and reusing offline optimization for real-time constrained control is a sensible direction.
  2. Deriving objective priorities from the measured physical state is intuitive and interpretable.
  3. The OPF results are promising: on IEEE 30-bus, GPC is within about 0.3% of the reference and close to MPC cost at 12 ms vs. about 1.2 s per decision.
  4. The experimental coverage is broad: three quite different domains (analytical control with convex and nonconvex objectives, static and moving multi-agent navigation, and AC optimal power flow under nominal, perturbed, and topology-shifted conditions), with feasibility, safety, optimality gap, and latency reported separately.
Critical weaknesses
  1. Missing key baselines and weak comparisons. GPC is essentially a learned solution map, but it is not compared with learning-to-optimize / OPF proxy methods (e.g., DC3), approximate MPC, or Pareto set learning. The RL baselines all reach 0% feasibility or safe success and are given less information than GPC, so the comparison is not very informative.

  2. Limited depth of evidence and no ablation. Despite the broad coverage, the main OPF comparison uses a single trajectory per setting, and the 118-bus study uses one test graph. There is no ablation of the main components (dual encoder, rollout-aware refinement, urgency-based priorities, Lie correction), and the offline solving and training cost is not reported.

Questions for authors
  1. How does GPC compare with a plain supervised solution map (e.g., an MLP from observation and priority to action) followed by the same projection and feasibility correction?
  2. What is the total offline cost, and after how many online decisions does GPC break even against MPC?
5 comments
AuthorsOpenReview ↗

1. How does GPC compare with a plain supervised solution map under the same projection and feasibility correction?

Reviewer comment — Questions For Authors 1:

How does GPC compare with a plain supervised solution map (e.g., an MLP from observation and priority to action) followed by the same projection and feasibility correction?

Authors’ response.

Under the same projection, safety mechanisms, and final refinement, both reach 100% success in moving scenes, and GPC attains the lower mean cost when priorities change externally (3988.25 versus 4045.59 for the supervised map).

New in the revision. No existing method performs amortized optimization for Pareto-optimal control with priorities that change online, so we built the closest amortized baseline ourselves. It is a supervised solution map trained on the same reference data and observable inputs as GPC and deployed with GPC's projection, safety mechanisms, and final candidate refinement. It does not include GPC's dual encoder, measurement-based localization, or latent updates.

Scene, prioritySuccess, GPC / supervised (%)Mean cost, GPC / supervisedMean time, GPC / supervised (ms)
Static, fixed90 / 952175.28 / 2146.36118.07 / 117.83
Static, external90 / 1004932.58 / 4858.54118.38 / 118.32
Moving, fixed100 / 1002402.11 / 2294.74121.55 / 121.28
Moving, external100 / 1003988.25 / 4045.59120.11 / 119.05

What GPC adds. Both methods amortize optimization, and this comparison measures closed-loop quality under identical final mechanisms. The supervised map returns one solution per input. GPC represents the response across priorities, so one trained controller follows changing priorities, and its localization, accepted updates, and feasibility margins let the validity of the response be checked as conditions change. New in the revision: the representation itself is tested separately; removing only the state-side representation worsens the Stage-I objective by 3.8210% (static) and 3.4929% (moving).

Evidence: Table 10 (p. 43); Appendix M.3 (p. 43); Table 9 (p. 43).

AuthorsOpenReview ↗

2. What is the total offline cost, and when does GPC break even against MPC?

Reviewer comment — Critical Weaknesses 2 (excerpt):

… and the offline solving and training cost is not reported.

Reviewer comment — Questions For Authors 2:

What is the total offline cost, and after how many online decisions does GPC break even against MPC?

Authors’ response.

Each hour of offline computation is recovered after about 3,050 online decisions against LRH-MPC. Appendix M.13 now reports the offline workload of every stage and the break-even formula.

In the submission. Algorithm 1 specifies the offline stages: context sampling, scalarized reference solves, and the two training stages. Table 2 reports the decision times: 1193.8 ms for LRH-MPC and 12.3 ms for GPC in Case30.

New in the revision. Appendix M.13 reports the offline workload and the break-even formula.

SceneReference labelsTraining-sample presentationsTraining batches
Static navigation60,0006.672M104,250
Moving navigation120,00012.864M201,000

The nominal OPF references cover 5000 loads × 50 weights, and the Case118 controller is trained once on 1,001 contingency graphs. For offline time ToffT_{\mathrm{off}} in hours, Nbreak−even=3600 Toff/(tMPC−tGPC)N_{\mathrm{break-even}}=3600\,T_{\mathrm{off}}/(t_{\mathrm{MPC}}-t_{\mathrm{GPC}}); with the Case30 times, each decision saves about 1.18 s.

Evidence: Appendix M.13 (p. 52); Algorithm 1 (p. 4); Table 2 (p. 9).

AuthorsOpenReview ↗

3. How does GPC compare with learning-to-optimize, approximate-MPC, Pareto-learning, and OPF baselines?

Reviewer comment — Critical Weaknesses 1:

Missing key baselines and weak comparisons. GPC is essentially a learned solution map, but it is not compared with learning-to-optimize / OPF proxy methods (e.g., DC3), approximate MPC, or Pareto set learning. The RL baselines all reach 0% feasibility or safe success and are given less information than GPC, so the comparison is not very informative.

Authors’ response.

GPC's reference gap is close to online optimization and below that of every learned baseline in Case30, and the Pareto-learning methods fail on 29–51% of analytical trajectories versus 2% for GPC.

In the submission. The main OPF comparison was against online optimization with perfect information and a state-of-the-art nonlinear solver (LRH-MPC with IPOPT), relative to which GPC's reference gap is small (Table 2); Table 2 also included TD3 and CPO, and Table 3 compared GPC with MPC, CPO, TD3, and online scalarization on the analytical tasks. To address the review, we added a baseline from each family it lists.

New in the revision.

  • Learning to optimize / OPF proxies: DC3, EqNN, DeepOPF-AC, DeepOPF-FT, and LD-ACOPF, together with NMO-OPF (IPOPT), which re-solves the nonlinear multi-objective AC-OPF at every step.
  • RL with matched information: TD3 and CPO-style PPO-Lagrangian re-run with GPC's per-step weights and an AC power-flow (AC-PF) correction.
  • Approximate MPC: a supervised solution map deployed with GPC's final mechanisms (Tables 10–11), and acados SQP/SQP-RTI, against which GPC reaches 90–100% safe joint success versus 0–20%.
  • Pareto-set learning: PPSL and explicit MOMPC on the analytical tasks, which fail on 51% / 29% and 41% / 33% of 100 trajectories (convex / nonconvex), versus 2% for GPC.
Case30 methodFeasible, nominal / perturbedGap (%), nominal / perturbed
GPC300/300 / 299/300+0.3011 / +0.2496
NMO-OPF (IPOPT)300/300 / 299/300+0.2752 / +0.2281
DC3300/300 / 297/300+0.3491 / +0.8198
EqNN300/300 / 200/300+0.3260 / +9.3347
DeepOPF-FT300/300 / 296/300+0.3148 / +0.7905
TD3 (AC-PF)300/300 / 299/300+0.5046 / +10.4191
CPO-style PPO-Lagrangian (AC-PF)300/300 / 298/300+5.5193 / +9.2208

Evidence: Table 14 (p. 46); Table 15 (p. 46); Table 3 (p. 10); Table 10 (p. 43); Table 13 (p. 45).

Comments on the paper

10 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers. Each reviewer's questions are answered point by point in that reviewer's own thread, and each answer is self-contained. Within each answer, In the submission marks evidence that was already in the submitted paper, and New in the revision marks experiments and text added in response to the reviews.

New in the revision. Matched amortized controls (a supervised solution map and DPC); OPF proxies (DC3, EqNN, DeepOPF-AC, DeepOPF-FT, LD-ACOPF) and NMO-OPF (IPOPT); RL re-run with GPC's per-step weights; acados SQP/SQP-RTI; PPSL, explicit MOMPC, and HoP on the analytical tasks; the full Case118 held-out test; component studies; and a numerical check of Theorem 1. They are in Appendix M (p. 42) and Appendix N (p. 53), with all new or materially revised text in blue. The main text still ends on page 10, and the abstract is unchanged. An anonymized evaluation package is included as supplementary material.

For easier reading, all of our responses are also collected in a single PDF, Response.pdf, in the supplementary material, with the same numbering as here.

AuthorsOpenReview ↗

General response to all reviewers

For easier reading, the detailed results are given point by point in each reviewer's thread and are collected in our response letter, Response.pdf, in the supplementary material, with the same numbering. Long reviewer comments are quoted as short excerpts so that each answer stays concise; the full reviews remain on this page.

Thank you for the comments. However, it feels increasingly difficult to assess this combined rebuttal comment (and the subsequent one). Can the authors take their time to respond to the individual reviewer queries point by point ?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.