acceptodds
Under review as a conference paper at ICLR 2027

Compact Closed-Loop Response Representations for Controller–Task Matching

Abstract

Reservoir computing enables robot controllers to exploit temporal information while training only an output readout. However, different trained reservoirs can interact differently with the same plant, making their suitability dependent on the requested motion. We introduce Closed-Loop Response Representations (CLRR) to compare trained nonlinear echo state network controllers through compact, reusable descriptions of their feedback behavior. For each plant, weighted principal component analysis learns a shared basis for local closed-loop response residuals from development controllers, without nonlinear performance labels. In the complex-response formulation, each finite-window multisine tracking task defines an analytical quadratic functional that evaluates the reconstructed response directly in its learned coordinates. After response characterization and encoding within a fixed operating regime, the same controller code supports new task queries without retraining or additional task-specific nonlinear rollouts. A distortion bound relates compression error to response-based task-cost error and controller-comparison margins. In near-equilibrium Cart-Pole simulations, eight coordinates achieve macro Kendall correlation 0.936 across 43 held-out controllers and 36 tasks, outperforming the tested compact spectral summaries. On a separate 16-controller test bank for a simulated seven-joint Panda arm, eight response coordinates plus seven equilibrium offsets achieve correlation 0.538, compared with 0.919 for the full response, yet select the best or second-best controller on all eight tasks. Full and sparse responses select the best throughout. These findings support compact response representations for task-dependent reservoir-controller comparison, while showing that useful controller choices can survive substantial losses in overall ranking fidelity. The quality of the representation therefore depends on the controller distinctions it preserves, not reconstruction accuracy alone.

Primary area:
learning on time series and dynamical systems
Keywords:
representation learning, closed-loop dynamics, controller–task compatibility, reservoir computing, nonlinear control, frequency response

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_uc95OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims to propose a method that is able to identify reusable descriptions of the feedback behavior of controllers based nonlinear echo state networks. This is supposed to enable the same controller to support new task queries without retraining. The method is evaluated on a cart-pole and robotic arm control tasks.

Critical strengths

None

Critical weaknesses
  1. The paper is extremely poorly written and badly presented. This seems to be mainly due to the fact that the most if not all content and text seems AI-generated without substantial revision by the authors. This makes it as a reader extremely hard to follow and, therefore, the paper does not provide any meaningful value to the research community even if the method could potentially be promising (the reviewer is not able to assess the research idea due to the poor presentation).
Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer uc95

Thank you for the review.

You ask whether we checked and revised the AI-assisted content. We have read the manuscript, its mathematical formulation, and the code ourselves, and we take responsibility for the work. AI assistance was disclosed in the submitted manuscript; the revised statement will further specify its scope and our checks.

We recognise that the paper was hard to follow. We are revising the presentation, including background before the method, a table defining the tasks, a consolidated arm-results section, and clearer, shorter explanations. We plan to upload the revised PDF by 18 November 2026.

Could you identify the passages that prevented you from assessing the method? Do your soundness assessment and ethics flag stem from the presentation and your concern about whether we checked the AI-assisted content, or have you identified additional problems with the mathematics, experiments, or reported results?

We will address these points directly and would appreciate a technical assessment of the method and evidence in the revised version.

Currently, the PDF does not show any highlighted changes to allow me to assess potential improvements in presentation and writing quality. Therefore, I will keep my rating of the paper.

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

The paper proposes Closed-Loop Response Representations (CLRR) to compare trained nonlinear echo state network controllers through compact, reusable descriptions of their feedback behaviour. For each controller, the closed loop is linearized around an equilibrium, and weighted principal component analysis, fitted on development controllers, compresses the resulting closed-loop frequency-response residuals into a low-dimensional code. An analytical multisine tracking functional for the compressed responses is designed to collect the necessary information for each controller, making the method suitable applied to different tasks. Simulations on Cart-Pole and Franka show that eight coordinates preserve strong ranking agreement on Cart-Pole, while on the arm the ranking

Critical strengths
  1. The experimental design and results honestly show both the advantages and the limitations of the proposed method.
  2. Theorical contribution and algorithm contribution are well reflected in the paper.
Critical weaknesses
  1. Presentation. The paper lacks a dedicated related-work section, and some basic background is very brief: for example, echo state networks and frequency responses are each introduced in only one or two sentences
  2. Lack of generalizability. The paper claims that the method applies to different tasks, but the tasks are not clearly presented. From the paper, I can only find that the task only differs in tracking trajectories.
  3. Unclear motivation for two variants. The paper introduces both H-CLRR and K-CLRR, but K-CLRR is evaluated only on Cart-Pole. It is unclear why both are needed, which may lead readers to question the necessity of each.
Questions for authors
  1. The paper introduces two variants, H-CLRR and K-CLRR, but K-CLRR applies only to scalar outputs and is evaluated only on Cart-Pole, while H-CLRR already covers both systems. What is the role of K-CLRR beyond Cart-Pole?
  2. Following this, can the method be applied to different systems or operating points without repeating the whole procedure?
  3. The paper compares several alternatives, but the necessity of the main components remains unclear. In particular, on the arm task, the method uses no learned basis, clearly outperforms the learned representation.
  4. Is the weighting S in the weighted PCA necessary. How does unweighted PCA perform?
3 comments
AuthorsOpenReview ↗

Response to Reviewer kWsQ

We thank the reviewer. We completed the tests below and are revising the manuscript.

Additional experiments. To check whether the best-or-second-best selections in the original manuscript would recur, we ran a repeat test on a new bank (group) of 64 trained ESN controllers. We followed the original training and evaluation procedure on eight original joint-tracking tasks and eight new tasks.

To address Q3 (the learned basis), we tested whether eight coordinates select better when controllers differ more in feedback behaviour. We created two more controller banks, Bank 1 and Bank 2. The original ESNs all imitated one feedback controller, called the teacher. In the new banks, each ESN imitated a teacher with different joint-feedback gains. We learned each basis from 96 controllers and tested on 64 separate controllers.

Bank 1 used the original Panda model. To test changed dynamics in Bank 2, we reduced servo gains tenfold and added a 2 kg payload. We designed Bank 1 after the repeat test and Bank 2 after Bank 1's results. Both reused the same 16 tasks and d = 8. We fixed the main predicted rankings before final simulations so the outcomes could not influence those rankings.

W1 — Related work and background. We are adding Related Work and Background sections that explain ESN controllers and closed-loop frequency responses before the method.

W2 — What the tasks are. We agree: the tasks vary the joint-angle trajectory, not the manipulation skill. The arm references combine two or three sinusoids within 0.1–1 Hz, with different frequencies, phases and joint amplitudes. For a new reference, we construct a quadratic cost function and apply it to the stored controller representations, without retraining controllers or running new nonlinear simulations. We did not evaluate other skills, constraints or non-periodic references. We will add a task table.

W3 and Q1 — Why K-CLRR? We included K-CLRR to test whether a simpler scalar squared-error description is sufficient. H-CLRR retains complex responses; the variants are alternatives, not successive stages. In the original Table 1, Cart-Pole Set B gives τ = 0.903 for H-CLRR versus 0.818 for K-CLRR; Sets A and C give similar correlations. We will present K-CLRR as a supporting scalar-tracking comparison, evaluated only on Cart-Pole. On the seven-joint arm, we retained complex responses to represent interactions between joints.

Q2 — Reuse on other systems or operating points. Partly. We must calculate each controller's response again, but we may be able to keep the learned basis. To test this, we reused Bank 1's basis at another joint configuration (posture) and on Bank 2's changed dynamics. H-CLRR selected the best of 64 newly trained controllers on 16/16 tasks at the second posture and on 15/16 in Bank 2. The posture test was mild: the nominal closed-loop response changed by 0.25%. Reuse did not always work: the original single-teacher basis gave low correlations on Banks 1 and 2 (τ = 0.19 and 0.26).

Q3 — What does the learned basis add? We learn the basis to reduce storage; response-based scoring does not require it. Sparse-eight instead stores responses at eight frequencies. On Banks 1 and 2, H-CLRR preserved nearly the same controller choices with about 41% less total storage. At 2–5 mrad amplitude:

Best controller selectedBank 1Bank 2
H-CLRR16/1615/16
Sparse-eight16/1616/16
One fixed ranking for all tasks3/168/16

The third row shows selection by one fixed ranking. H-CLRR stores 15 values per controller: 8 coordinates and 7 steady-state offsets. Sparse-eight stores 791: eight 7 × 7 complex response matrices and 7 offsets. Including the basis and stored task functions, totals are 36,208 versus 61,706 for 64 controllers and 16 tasks. On Bank 2's remaining task, H-CLRR selected the second-best controller. Across all 16 tasks, the average extra cost above the best controller was 0.074%. Sparse-eight still ranked better (τ = 0.995 versus 0.967; 0.976 versus 0.907).

We agree that sparse-eight performed better in both tests with one teacher. In the repeat test, sparse-eight selected the best controller on 4/8 original and 6/8 new tasks; H-CLRR achieved 1/8 and 0/8. We will retain this result alongside Banks 1 and 2.

Q4 — Is weighting S necessary? No. As requested, we repeated PCA fitting without weighting, keeping the controllers and tasks unchanged. Both versions selected identical controllers in every arm test. On Cart-Pole, averaging costs over valid trials, unweighted PCA gave τ = 0.927 versus 0.936, the same best-selection count (34/36), and lower average cost above the best controller (0.11% versus 1.81%). The quadratic cost identity needs no special weighting. We will describe S as optional and report the comparison.

We plan to upload the revised PDF before the discussion ends, separating the new experiments from the original results. We welcome the reviewer's assessment of these results for Q3.

I thank the authors for their detailed response and additional experiments. The new results partially address my concerns, particularly regarding the role of the learned basis and its storage efficiency. However, I still have reservations about the overall novelty and generalizability. I will maintain my original score.

We thank the reviewer for the follow-up.

Novelty. Our contribution is a compact description of each trained plant–controller loop from which we calculate predicted tracking costs for different references. Each reference defines an analytical quadratic cost on the stored coordinates, accounting for phase and interactions between joints. This requires no performance predictor fitted to nonlinear task outcomes. We also analyse how compression affects controller comparisons and test whether the compact descriptions preserve useful selections on controllers excluded from representation fitting.

Generalisability. CLRR supports generalisation across tracking references by reusing the same controller description for different movement speeds, amplitudes and coordination between joints. This reuse follows from the frequency-response formulation. The formulation permits richer sinusoidal combinations within the characterised frequency band and local operating regime; our experiments evaluate two- and three-component references.

We also demonstrated reuse across new controllers and changed arm dynamics. We applied the shared response basis learned from Bank 1 to Bank 2, where we reduced servo stiffness and damping tenfold and added a 2 kg payload. We calculated the new controllers’ responses and encoded them using the unchanged Bank 1 mean and basis. H-CLRR selected the best controller on 15 of the same 16 references at the original amplitude, without refitting the basis.

Reviewer_2oVqOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies controller-task matching for banks of already-trained nonlinear echo state network (ESN) controllers. Rather than running every controller on every new task, the authors characterize each plant-controller closed loop by its local complex frequency response and compress response residuals using weighted PCA fitted on development controllers. H-CLRR then evaluates finite-window multisine tracking tasks through an analytical quadratic cost in the compact code; a scalar-kernel variant (K-CLRR) is also studied for Cart-Pole. On 43 held-out Cart-Pole controllers across 36 tasks, an eight-coordinate code achieves macro Kendall correlation 0.936. On a simulated seven-joint Panda arm, eight coordinates plus seven offsets achieve correlation 0.538 versus 0.919 for the full response, but select the best controller on five tasks and the second-best on three.

Critical strengths
  • The paper introduces a compact and reusable representation for task-dependent controller selection. By combining weighted PCA of closed-loop responses with an analytical quadratic task-cost functional, the proposed CLRR framework enables efficient evaluation of new tasks without retraining controllers or performing additional task-specific nonlinear rollouts. The accompanying distortion bound provides theoretical support for the relationship between compression error and controller comparisons.
  • The experimental results demonstrate that compact representations can preserve useful controller-selection information. On Cart-Pole, the eight-dimensional H-CLRR achieves a Kendall correlation of 0.936 across 43 held-out controllers and 36 tasks, outperforming the tested compact spectral baselines. On the Panda arm, despite a lower overall ranking correlation, the method selects the best or second-best controller on all eight tasks. These results support the practical value of the approach, particularly when identifying high-performing controllers matters more than recovering the complete ranking.
Critical weaknesses
  • The evaluation is restricted to narrow local operating regimes, limiting evidence of generalization. The Panda arm tests use one contact-free posture, 0.1–1 Hz responses, and only 2–5 mrad excursions; all candidates are pre-screened for local stability and viability.
  • The practical advantage over sparse frequency sampling remains uncertain on the Panda arm. Sparse-eight has higher ranking correlation (0.925 versus 0.538), selects the best controller on all eight tasks, and uses less total descriptor storage for 16 candidates (12,664 versus 33,568). Preparing the latent task functionals also takes about 174 ms.
  • The eight-coordinate PCA code loses substantial controller-ranking information on the Panda arm. Increasing the development bank from 24 to 96 improves Kendall correlation by only 0.052, with a confidence interval including zero. Moreover, the distortion bound covers compression within the local-response model, not the approximation error relative to nonlinear rollouts.
Questions for authors
  • Could the authors provide an end-to-end, amortized storage and runtime comparison with sparse-eight as the number of controllers and repeated task queries grows, including response characterization and task-functional setup?
  • Could the authors analyze which discarded PCA directions cause arm ranking reversals, or compare with task-aware or ranking-aware compression? How well do the observed reversals match the proposed distortion/margin analysis?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 2oVq

We thank the reviewer. We completed the tests below and are revising the paper.

Additional experiments. To check whether the original best-or-second-best selections would recur, we ran a repeat test with 64 new ESN controllers. We followed the original procedure, using the eight original joint-tracking tasks and eight new tasks.

For W2 and W3, we tested whether eight coordinates select better with more varied feedback behaviour. The original ESNs imitated one feedback controller, called the teacher. In two new controller banks/groups (Banks 1 and 2), each ESN imitated a teacher with different joint-feedback gains. We fitted each basis on 96 controllers and tested 64 others. Bank 1 used the original Panda. To test changed dynamics in Bank 2, we reduced servo gains tenfold and added a 2 kg payload.

We designed Bank 1 after the repeat test and Bank 2 after Bank 1's results. Both reused the 16 tasks and d = 8. We fixed the main predicted rankings before final simulations so the outcomes could not influence those rankings.

W1 — Narrow operating regime. To address this concern, we tested tenfold larger movements (20–50 mrad) and another joint configuration (posture). Sparse-eight stores responses at eight frequencies without learning a basis. Best-controller selections out of 16 tasks:

BankAmplitudeH-CLRRSparse-eight
12–5 mrad16/1616/16
22–5 mrad15/1616/16
120–50 mrad15/1614/16
220–50 mrad12/1614/16

At 20–50 mrad on Bank 2, H-CLRR's choices cost 1.57% more than the best controller on average and 18.45% more on the worst task (sparse-eight: 0.35% and 5.36%).

At the second posture, H-CLRR reused Bank 1's basis and selected the best of 64 newly trained controllers on 16/16 tasks. This was mild: the nominal closed-loop response changed by 0.25%. All tests remain simulated, contact-free, within 0.1–1 Hz, with controllers screened for local stability and viability.

W2 and Q1 — Benefit over sparse sampling.

At 2–5 mrad on Banks 1 and 2, H-CLRR preserved nearly the same choices with about 41% less total storage. One fixed ranking selected the best controller on 3/16 and 8/16 tasks. Sparse-eight ranked better (τ = 0.995 versus 0.967; 0.976 versus 0.907).

In the repeat test with one teacher, H-CLRR selected the best controller on 1/8 original and 0/8 new tasks; sparse-eight achieved 4/8 and 6/8. The best-or-second-best result did not recur. We retain the repeat test alongside Banks 1 and 2.

For Q1, we optimised and timed all three implementations. The earlier implementations made H-CLRR's set-up slow (174 ms) and overstated the query-speed advantage. Scores and controller choices did not change. For 64 controllers and 16 tasks:

MeasurementH-CLRRSparse-eightFull response
Set-up plus scoring (ms)1.042.360.74
Scoring prepared tasks (ms)0.261.050.39
Values, with shared arrays and task functions36,20861,706204,908

Timing excludes training, screening, shared reference-signal preparation, response calculations and basis fitting. H-CLRR stores 15 values per controller (8 coordinates + 7 steady-state offsets), versus 791 for sparse-eight (eight 7 × 7 complex matrices + 7 offsets). With the basis and these 16 task functions fixed, H-CLRR stores less from 32 controllers; sparse-eight stores less at 16.

H-CLRR was faster than sparse-eight; the full response was fastest with set-up. We claim no end-to-end speedup. With the older implementation, going from 16 trained controllers to scores took about 2.7 s for H-CLRR or sparse-eight, mostly response calculations. Response calculations and basis fitting for 96 controllers took 16.1 s.

W3 and Q2 — Compression and the bound. We agree for the original arm test: τ = 0.538 versus 0.919 for the full response. The gain from 24 to 96 fitting controllers remains inconclusive. Banks 1 and 2 lost less information (τ = 0.967 and 0.907, versus 0.996 and 0.978).

For Q2, we restored discarded directions in the original arm test. H-CLRR reversed 221/960 pair orders relative to the full response. Restoring directions 9–16, 17–32 or 33–95 separately repaired 70, 98 and 90 but introduced 47, 55 and 48 reversals. Restoring all repaired 162, introduced 54 and raised τ to 0.750. No block fixed every comparison.

H-CLRR preserved the order for 53% of pairs with full-response cost gaps below 0.1 times the task's cost standard deviation and 94% with gaps at least twice that deviation. Close comparisons were more vulnerable, as the margin analysis suggests. Proposition 1's condition certifies none of the 960 pairs, although H-CLRR preserves 739 orders. The bound covers compression, not local-model error, and is too conservative here.

Tests on fitting controllers favoured the original over the task-derived PCA weighting. We did not train the basis to minimise ranking errors.

We plan to upload the revision before discussion ends, reporting new results separately. We welcome feedback on W2 and Q1.

I appreciate the additional results provided in the rebuttal, which resolved some of my concerns.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers and the Area Chair. The individual replies give the details.

The problem. We have a bank of trained echo state network controllers. For each new tracking task, we want to choose the best controller without test-running every controller on it. After calculating the local robot–controller response, H-CLRR stores 15 numbers per controller on the seven-joint arm. They summarise how the robot responds with that controller in the loop, and H-CLRR estimates the tracking cost from them. We ask whether so few numbers preserve the right choice.

What changed in our claim. The original manuscript reported that H-CLRR selected the best or second-best controller on all eight arm tasks. In a repeat test with 64 new controllers, H-CLRR selected the best controller on only 1/8 original and 0/8 new tasks. We no longer present that result as general.

We then tested two new banks whose controllers differ more in feedback behaviour (Banks 1 and 2), on the same 16 tasks as the repeat test. We compared H-CLRR with sparse-eight, which stores uncompressed responses at eight frequencies. For 64 controllers and 16 tasks, H-CLRR preserved nearly the same controller choices as sparse-eight with about 41% less total storage, including shared arrays and task functions. At the original amplitude, H-CLRR selected the best controller on 16/16 and 15/16 tasks (sparse-eight: 16/16 in both). At tenfold larger movements, H-CLRR reached 12/16 on Bank 2 (sparse-eight: 14/16). Sparse-eight ranked the controllers more accurately at the original amplitude on both banks. We claim storage savings, not an end-to-end speedup. All tests are simulations.

Corrections. The original timing comparison used unequally optimised implementations and overstated H-CLRR's query-speed advantage. We optimised all implementations and repeated the comparison (reply to 2oVq, Q1). Scores and controller choices did not change. We also made the storage accounting consistent across methods. The reproducibility statement wrongly said that a source package accompanied the manuscript. We uploaded only the PDF and are correcting this.

Revision. We plan to upload the revised PDF before discussion ends. We keep the method, the theory and the original experiments, and add background, a task table, a section for the new tests and an expanded AI-use statement.

AuthorsOpenReview ↗

Revised manuscript uploaded

We have uploaded the revised manuscript. Changes are coloured by reviewer: blue for 2oVq, rust for kWsQ, and green for uc95 and presentation. The main changes are:

  • Section 2 now contains related work (2.1), background on reservoir controllers and the closed loop (2.2), and the problem setting with a table that defines a task (2.3).
  • Sections 5.4 and 7 report the repeat test with 64 controllers, Banks 1 and 2, the tenfold larger movements and the second posture. Appendix K gives the protocols and the full results.
  • Section 6.5 counts storage in the same way for all methods and times equally optimised implementations.
  • Section 4 states what the bound covers and reports the margin analysis.
  • The abstract, the introduction and the conclusion state the revised claim. The AI-use and reproducibility statements are updated.

We thank the reviewers and the Area Chair for their time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.