acceptodds
Under review as a conference paper at ICLR 2027

Equivariant Counterfactual Supervision for Hidden-Dynamics Representation Learning

Abstract

We introduce equivariant counterfactual supervision for learning physical context from a robot's interaction history. The method compares simulated responses under controlled changes to an environmental property, using comparison rules appropriate to material and geometric changes. Subtracting the mean response across the compared conditions produces residual targets. A reservoir feature bank summarizes the interaction history, and a learned code predicts these targets at a query from another rollout in the same environment. We evaluate what physical information the code makes accessible and how useful it is to new prediction tasks. Relative to Full-transition supervision, Residual raises steepness linear-probe from 0.555 to 0.907 and improves mean exact-cell retrieval in all ten factor-level holdouts. A paired target-centering ablation corroborates the accessibility difference under matched heads and scales. Fresh predictors show a modest in-distribution full-transition benefit, whereas Full has lower mean error on all four held-out factor-response tasks. The results establish a supervision-dependent accessibility effect whose predictive value depends on the receiving task.

Primary area:
learning on time series and dynamical systems
Keywords:
hidden dynamics, counterfactual supervision, equivariant learning, contact dynamics, representation learning, quadruped locomotion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_mHqYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

A legged robot's response to an action depends on physical properties it cannot measure directly, such as terrain slope, friction and ground softness, so it must infer them from its recent interaction history. Prior history-based adaptation methods typically train the history encoder to predict the robot's next state. The authors argue that good next-step accuracy does not tell us whether the encoder has actually captured these properties. They propose an alternative training signal, Equivariant Counterfactual Residual (ECR) supervision. In simulation, they replay the same state and action while changing one physical property. They subtract the average response across those variations, leaving targets that contain only the differences caused by the property. A compact code computed from a history taken from a different trajectory in the same environment is then trained to predict these difference targets. The history is first summarized by a fixed random recurrent network, followed by a small trained network.

The experiments use a simulated quadruped across a grid of 60 slope/friction/softness combinations. ECR is compared against next-step prediction and against adapted versions of several prior objectives, including one trained directly on the true property labels. ECR codes make slope far easier to read out than next-step codes do, and they group histories from property values held out of training more faithfully. A matched ablation suggests the subtraction step contributes to this. However, when the frozen codes are used to train new predictors, the in-distribution gain is small. Next-step prediction wins on every held-out prediction task, and at unseen slopes ECR is worse than using no history code at all. The authors conclude that the supervision makes physical factors easier to read out, but that whether this helps prediction depends on the downstream task.

Critical strengths
  1. Slope is consistently easier to read out of ECR codes than out of next-step-prediction codes.
    • The advantage over next-step prediction (NSP) and CaDM-style training appears in three different measurements: mutual information, linear probes and nearest-neighbour retrieval (Fig. 2).
    • It holds when only one time step per trajectory is used, so repeated near-duplicate samples do not inflate it (Table 10: normalized mutual information 0.900 vs 0.148).
    • It holds in all ten tests with a property value held out of training (Tables 11-13).
    • Swapping in a history from a different slope raises ECR's slope-prediction error 1.60x but leaves NSP's unchanged (1.00x), so the ECR predictor does use the slope information (Table 8).
  2. The paper reports its negative results plainly. It states that NSP has lower error on all four held-out prediction tasks (Table 3). The discussion (Sec. 6) admits that the ablation does not isolate the effect of subtracting the mean. This makes the paper a useful counterexample: a code from which physical factors are easy to read out need not help a predictor.
Critical weaknesses
  1. Easier read-out of the factors does not improve prediction, and at unseen slopes the ECR code makes prediction worse than having no code.
    • At the held-out 5° and 10° slopes, a predictor given the ECR code has 15% and 18% higher error than one given only the current state (0.1601 vs 0.1394 and 2.4578 vs 2.0897). NSP codes beat both (Table 3, Tables 14 and 17).
    • In distribution, ECR's next-step error is only 1.8% lower than NSP's (App. H.3).
    • The paper therefore shows no prediction task where easier read-out pays off, and it does not test control. The two baselines that also use environment information are LDG-style (groups histories from the same environment) and RMA-style (trained on the true labels). Both match or beat ECR on slope read-out (linear-probe R² 0.940 and 0.977 vs 0.907; Sec. 5.1).
  2. The evidence does not isolate what is new about ECR. The ablation meant to isolate mean-subtraction is confounded, NSP receives far less supervision than ECR, and the step beyond the friction-only predecessor is never tested.
    • The control trains on the same targets with the mean left in. In some coordinates these targets are up to about 340 times larger, so the gradient is clipped on every update, with average pre-clipping norms above 3,100 (Table 25, App. H.3).
    • The control ends up worse on slope than ordinary NSP (linear-probe R² 0.380 vs 0.555; Table 1). The paired gain may therefore come from a badly scaled baseline rather than from removing the shared component. The authors acknowledge this (Sec. 5.2.1), yet the abstract says the ablation "corroborates" the effect.
    • ECR is trained to predict how the response changes when exactly the three factors (slope, friction, contact type) are varied. That uses simulator replays NSP never sees, and the probes and retrieval tests then try to recover those same factors. A code trained this way is expected to expose the factors more readably than a next-step code. The gap over NSP therefore shows that the extra supervision is used, not that the residual construction is a better learning principle.
    • The paper presents itself as extending an earlier friction-only version of the method to slope and multiple factors (Sec. 2). The new ingredient, a replay that keeps the robot's position on the slope comparable across conditions, is never compared with a simpler replay, so the contribution beyond that predecessor is not shown.
  3. The results are hard to follow because the main text keeps recording where each number came from. Sentences and captions interrupt the findings with run lineage, reused encoders and sample counts (e.g. "Fresh 5000-update heads on the existing NSP and CaDM encoders provide context; their encoder training is not paired with the new ECR run", Sec. 5). Several terms are used without definition, such as "donor history", "receiving query", "anchor" and "accessibility". The title also promises an "equivariant" method, although the network is not equivariant (Sec. 2); equivariance is only a property checked in the simulator (Sec. 3.1).
Questions for authors
  1. (Weakness 1: no prediction benefit) Is there any downstream use where the ECR code beats both next-step-prediction codes and a predictor given no code, for example closed-loop locomotion or adaptation at the held-out slopes? At those slopes the ECR code currently makes prediction worse than no code (Tables 14 and 17). If the code helped a controller there, the main weakness would be largely resolved. If it hurts control as well, the claim that the code captures useful physical context does not hold.
  2. (Weakness 2: confounded ablation) Does subtracting the mean still help once the control is properly scaled? The control trains on the same replayed responses without the subtraction. It could, for example, standardize its targets with their own statistics, or have its clipping and learning rate tuned so it is not clipped on every update. If a control fixed this way matches ECR's slope read-out, the gain comes from the replay data rather than from the residual construction. The contribution would then reduce to "train on counterfactual replays".
  3. (Weakness 2: extra supervision) Why are counterfactual replays needed at all? The LDG-style baseline only needs to know which histories come from the same environment, and the RMA-style baseline uses the true labels. Both match or beat ECR on slope read-out, and RMA-style also beats it on retrieving the exact environment. Showing ECR ahead on some measure, at equal environment information and simulation cost, would justify re-simulating every factor setting. Without that, ECR is not the best choice among methods with similar access to the simulator.
  4. (Weakness 2: what is new beyond the friction-only predecessor) What does this paper establish that the earlier friction-only version of the method, which it says it extends, did not? If the new part is mainly the extension to slope and multiple factors, the paper should show that its matched replay for slopes matters, for example against a simpler replay that does not keep the robot's position on the terrain comparable. Without that, the contribution over the predecessor is small.
3 comments
AuthorsOpenReview ↗

Response to Reviewer mHqY

We thank the reviewer. Numbered section and appendix references refer to the original manuscript.

W1, Q1. Does the representation improve a downstream task?

Yes, for next-step prediction in training conditions. To answer the reviewer's Q1, we trained new predictors from frozen representations. We added a trainable transformer to test dependence on the fixed reservoir, with the same encoder architecture across objectives.

With 6,000 encoder updates and 3,000 predictor updates, mean normalized prediction MSE was:

History representation added to the state–action queryMSE
ECR0.0305
NSP0.0315
CaDM-style0.0316
None (Query-only)0.0364

Representation results average three encoder seeds; the last row averages two predictor seeds.

We also asked whether a steepness predictor learned in one condition works in another. We fitted regularized linear predictors on source-condition training rollouts and evaluated target-condition test rollouts: friction 1.2↔0.2 and Rigid↔Soft+Dissipative contact profiles. Steepness is absolute slope. All conditions were included in encoder training and training-data preprocessing.

We added NSP trained only on recorded transitions, without counterfactual replays (factual-only NSP). Ordinary read-out fits a predictor across conditions; transfer fits in one and tests in another. Reservoir results average three encoder seeds; transfer averages four tests at regularization strength λ=100 (Spearman correlation).

RepresentationOrdinary steepness R²Transferred correlation
ECR0.9070.918
Factual-only NSP0.8810.128
NSP0.5550.282
LDG-style0.9400.233
RMA-style0.9770.925

Similar ordinary R² masks a substantial transfer difference. Across six regularization strengths, ECR's correlation spans 0.74–0.92 versus factual-only NSP's 0.11–0.19. ECR's mean transferred numerical R² is 0.71 at λ=100 and negative at every tested λ≤10.

The transfer test has no Query-only reference. Factual-only NSP slightly overtakes ECR in prediction after 20,000 encoder updates. NSP beats ECR on all four held-out prediction tasks; ECR loses to Query-only at both slopes. The stop-line control test did not establish a benefit: ECR minus no representation is +0.05 reward, with a paired 95% t interval [−2.58, 2.67] over five policy seeds using three representation seeds.

W2, Q2. Does the gain survive proper target normalization?

Yes. In the original paired comparison, ECR-Paired learns response differences and NSP-Matched learns full responses with the same training decoder architecture. We tested the reviewer's scaling and clipping repairs, keeping the remaining setup fixed.

Paired comparisonSteepness R²
ECR-Paired0.900
NSP-Matched with its own target standardization0.450
NSP-Matched without clipping0.423

These are three-run means. The standardized control clipped fewer than 0.6% of updates. Standard NSP is separate and already standardizes its own targets (App. B.2).

The original decoder is not explicitly given slope direction. ECR's advantage holds in this tested setting with partially observed inputs. Supplying direction to the reservoir decoder let NSP catch up: 0.970 versus ECR's 0.968. NSP also caught up after 20,000 transformer updates when replays kept slope direction consistent. ECR requires neither change.

W2, Q3. What do replays add beyond NSP, LDG-style and RMA-style?

NSP already received the same counterfactual data as ECR (§3.5; Apps. B.2–B.3). The history encoder architecture is shared; objectives retain their own output networks and normalization.

In transfer, ECR exceeds LDG-style's environment-grouping supervision and approaches RMA-style's direct physical-label supervision. ECR uses response differences, intervention metadata and histories paired by environment; slope replay also uses geometry. Computing its representation targets does not require numerical factor labels.

W2, Q4. What is new beyond the friction-only method?

The predecessor used residual supervision with matched friction interventions and histories from different rollouts. We extend this to slope and contact properties. For slope, we transform robot and terrain together to preserve their relative configuration while gravity remains fixed (§3.1). We have not completed the reviewer's proposed comparison with replay that tilts the terrain while leaving the robot fixed.

W3. How will we clarify presentation and equivariance?

We will add these results, revise the abstract's ablation claim, define the four flagged terms and move run details to the appendix. We will explain that the equivariance check rotates the robot, terrain and gravity together; the encoder is not an equivariant architecture.

Thank you for the new experiments. The transfer results are interesting. However, the gain in next-step prediction is very small, standard next-step prediction catches up once slope direction is available, and the comparison with the friction-only predecessor is still missing. My assessment therefore remains the same.

Thank you for the follow-up.

We do not claim that ECR beats NSP on next-step prediction. Our question is whether a history representation can retain the information needed to predict the robot’s next motion while allowing the same steepness read-out to work across different surfaces. Since NSP trains directly on the next physical response, it is the natural baseline for testing whether ECR preserves that predictive ability. Requiring ECR to outperform NSP on one-step prediction would test a different claim.

Toso, LeCun, Anderson, and Bounou (2026, Lemma 2) illustrate why both evaluations matter. In their linear JEPA example, an encoder can discard every unstable mode while predicting its next latent state with zero error. In our experiment, the same transformer architecture gave similar next-step errors on new rollouts from familiar conditions (0.0305 for ECR vs. 0.0315 for NSP). Yet when a steepness read-out fitted in one condition was applied in another, mean Spearman correlation across four friction and contact transfers was 0.878 for ECR vs. 0.099 for NSP at λ = 100.

With the original replays, inputs, and training budgets we tested, NSP cannot catch up to ECR on steepness read-out. Neither the encoder nor the training decoder explicitly received the direction of the replayed slope. After 20,000 transformer updates, steepness R² was 0.897 for ECR vs. 0.279 for NSP.

Our analysis in §3.4 explains why we tested direction. Uphill and downhill replays can give the decoder nearly identical terrain-relative queries while gravity produces different physical responses. We therefore gave only the training decoder a privileged two-dimensional vector indicating downhill direction relative to the robot’s heading. A probe given the query and this vector distinguished 5° from 10° at chance (50%). With the added direction input, reservoir NSP caught up (R² of 0.970 for NSP vs. 0.968 for ECR). For the transformer, ECR remained ahead at 6,000 updates (0.975 vs. 0.841). The transformer tie at 20,000 updates arose in a separate experiment that changed the replays to keep slope direction consistent (0.988 for NSP vs. 0.987 for ECR). Here, the geometry let us identify the missing variable and the simulator let us supply it. In another partially observed system, either step may be difficult.

Beyond the friction-only predecessor, we introduce matched slope replays and test whether steepness can be read across changes in friction and contact. Varying friction leaves the supporting terrain in place. For slope, we rotate the robot and terrain together so the feet begin in the same configuration relative to the ground while gravity remains fixed. The direct comparisons with friction-only training and terrain-only tilt remain open. The result we establish is that, under the matched replays, a steepness read-out fitted on one surface continues to rank slopes on another.

Reference: Toso, LeCun, Anderson, and Bounou (2026). Preserving Unstable Modes Through Inverse Dynamics in JEPA World Models, Lemma 2.

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

This paper proposes a method for learning representations of hidden environmental variables from robot trajectories through controlled interventions. The method first constructs counterfactual transitions by changing one environment variable at a time, and computing a residual target. The representations are then trained to predict these residual targets given their corresponding state-action queries.

The authors conduct experiments to evaluate the represenatation on multiple metrics including the environment variables' accessibility from the representations, a retrieval-based metric, and dynamics prediction error. No downstream usage of the representation is evaluated. And no real robot hardware is used.

Overall, the paper shows that counterfactual residual supervision can make certain physical properties more accessible in learned representations, but the contribution is minor and it does not translate to better downstream perfomance of the robot. I therefore recommend rejection of the paper.

Critical strengths

Learning representations via controlled intervention is a promising direction for disentangling confoudning factors that affect the representation. This paper offers a way to learn such a representation, although it requires privileged simulator access to reset the robot to the same state and evaluate multiple counterfactual transitions.

The different metrics used in the experiments shed lights on the properties of the representation. The paper reports negative downstream results rather than relying solely on favorable representation metrics.

Critical weaknesses

Limited novelty and practical application of the method. Prior works such as RMA have already demonstrated that representations learned from interaction histories can capture hidden dynamics and achieve generalization under changing dynamics in the real world. This paper proposes an alternative method to learn such a representation but does not show any downstream results or real world experiments. This makes the contribution of the paper very marginal.

This method relies on privileged access to a simulator that requires resetting the simulator to arbitrary states for generating counterfactual futures. However, prior works like RMA do not require this. This makes the experiment setting different from prior works, and one would expect to see significant gains in terms of downstream performance with this method, which is currently missing from the paper.

Questions for authors

How does this method compare to prior works when applied to the downstream control?

1 comment
AuthorsOpenReview ↗

Response to Reviewer iiKy

We thank the reviewer. Numbered section and appendix references refer to the original manuscript.

W1. What does ECR add beyond RMA?

ECR learns physical context from matched response differences without regressing numerical physical-factor labels. The new transfer test shows why this supervision matters: a steepness read-out fitted to ECR representations remains useful across the tested changes in friction and contact profile. Our contribution is representation supervision, with control as a downstream evaluation.

For the reviewer's RMA comparison, we use a label-supervised reference with the same history encoder architecture. Both original RMA and our RMA-style baseline use ground-truth environment parameters during training. RMA trains its history module to estimate a privileged latent derived from those parameters; our baseline directly predicts the factor values.

We tested transfer by fitting regularized linear predictors on source-condition training rollouts and evaluating target-condition test rollouts: friction 1.2↔0.2 and Rigid↔Soft+Dissipative contact profiles. Steepness is absolute slope. Representations were frozen; all conditions were present during encoder training and training-data preprocessing.

We added NSP trained only on recorded transitions, without counterfactual replay (factual-only NSP). Reservoir results average four transfers and three encoder seeds, with regularization strength λ=100.

RepresentationTransferred Spearman correlation
ECR0.918
RMA-style, trained on factor labels0.925
NSP0.282
Factual-only NSP0.128
CaDM-style0.437
LDG-style0.233

ECR approaches the label-supervised reference on this test. ECR and factual-only NSP have close ordinary steepness R² (0.907 versus 0.881), yet sharply different transfer. Across six regularization strengths, ECR's correlation spans 0.74–0.92 versus factual-only NSP's 0.11–0.19; RMA-style reaches 0.89–0.92. ECR's mean transferred numerical R² is 0.71 at λ=100 and negative at every tested λ≤10.

W1, W2 and Question. Does ECR improve downstream prediction or control?

For prediction within training conditions, yes. For control, we have not established an advantage.

The original manuscript evaluated new predictors trained on frozen representations (§5.2). To test dependence on the fixed reservoir, we added a trainable transformer encoder architecture shared across objectives. With 6,000 encoder updates and 3,000 predictor updates, ECR had the lowest mean normalized next-step MSE among the original objectives: 0.0305 versus RMA-style's 0.0312 across three encoder seeds. Factual-only NSP slightly overtook ECR after longer training. NSP has lower mean error than ECR on all four original held-out prediction tasks.

To answer the reviewer's control question, we trained PPO policies with frozen representations on a simulated Go1 stop-at-line task, using five policy seeds and three representation seeds.

Policy's history representationMean post-signal reward
ECR representation4.43
RMA-style representation4.98
No history representation4.38

ECR minus no representation was +0.05, with a paired 95% t interval [−2.58, 2.67]. ECR minus RMA-style was −0.55 [−2.71, 1.61]. These intervals over policy seeds establish neither superiority nor equivalence. This compares representation supervision, not the complete RMA controller.

W2. Why does ECR require simulator resets?

To address the reviewer's concern about simulator access, resets let us measure how the same state–action pair responds under changed physical conditions. Those response differences provide ECR's training targets.

Simulation provides matched replays for training and known factor values for evaluating physical read-outs. These controls are difficult to reproduce on hardware. RMA also trains in simulation and evaluates its adaptation-and-control system on hardware. Our experiments evaluate representation supervision; hardware transfer remains untested.

Each response stays linked to its state–action query and a history from that response's physical condition. ECR uses intervention metadata and, for slope replays, geometry.

NSP in the original manuscript received the same replay data (§3.5; Apps. B.2–B.3). The factual-only baseline separately tests training without those replays. Together, these comparisons support ECR's supervision design; they do not isolate the effect of resets alone.

We will add the transfer and control results and clarify the information each objective requires, emphasizing the physical information learned from response contrasts.

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

This study proposes Equivariant Counterfactual Residual (ECR) supervision for learning a context representation (code) from a robot's interaction history. The context is some physical parameter, such as terrian slope, friction coefficient, contact softness etc. that is unobserved but remains fixed within each rollout. The method varies the parameters in simulation and subtracts the mean response across the compared conditions. A history encoder and prediction network are then trained to predict the residuals using history from a different rollout under identical environmental conditions. In simulated quadruped experiments, ECR's code outperforms next step prediction (NSP) making it easier to estimate terrain steepness and identify histories collected under the same physical conditions. These gains however do not translate consistently into better prediction under unseen physical conditions.

Critical strengths

The improvement in recovering terrain steepness is substantial and is supported by several checks including linear probes, retrieval and history controls. The authors also transparently report where their method performs unfavorably. The finding that easier recovery of the physical parameter need not necessarily improve prediction is useful.

Critical weaknesses
  1. The main comparison does not clearly establish why ECR helps. Subtracting the mean response also scales down prediction targets and alters training behavior substantially. It remains unclear whether ECR's advantages would persist against a baseline with suitable normalization and training settings.

  2. The added value of ECR beyond recovering physical properties appears limited. Standard NSP performs better on all four heldout prediction tasks. The main experiments also use only one robot and one fixed reservoir and the broader utility of the method remains insufficiently demonstrated.

Questions for authors
  1. What does the 79 dim history token contain? How are the robot trajectories collected?

  2. Does the accessibility improvement remain when the uncentered baseline uses suitable target normalization and training settings?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Vxwo

We thank the reviewer. Numbered section and appendix references refer to the original manuscript.

W1, Q2. Does the gain survive suitable normalization and training settings?

Yes. In the original paired comparison, ECR-Paired learns response differences and NSP-Matched learns full responses with the same training decoder architecture. To test the reviewer's scaling explanation, we changed target normalization or gradient clipping, keeping the remaining setup fixed.

Paired comparisonSteepness R², three-run mean
ECR-Paired0.900
NSP-Matched with its own target standardization0.450
NSP-Matched without clipping0.423

The standardized control clipped fewer than 0.6% of updates. Thus, these repairs do not recover ECR's steepness read-out. Steepness means absolute slope. Standard NSP already used its own target normalization (§3.5; App. B.2).

To address the reviewer's concern about the fixed reservoir, we added a fully trainable transformer history encoder. Each objective used the same transformer encoder architecture. Across four learning rates at 6,000 updates, NSP reached steepness R² 0.21–0.26 versus ECR's 0.80–0.92. At 20,000 updates, NSP reached 0.279 and ECR 0.897. The advantage is therefore not confined to a fixed reservoir.

W1. What explains the advantage?

ECR subtracts the mean response across matched interventions, making response differences the prediction target.

We investigated the reviewer's mechanism question by changing the information available to the training decoder. In the original experiment, the decoder is not explicitly given the direction of the replayed slope. When we supplied direction to the reservoir decoder, NSP caught up: R² 0.970 versus ECR's 0.968. Changing the replays to keep slope direction consistent also allowed transformer NSP to catch up after 20,000 updates.

This supports ECR under the original, partially observed inputs. Additional direction information or a changed replay construction helps NSP catch up.

W2. Does the representation help beyond ordinary physical-factor read-out?

To address the reviewer's question about added value, we fitted regularized linear steepness predictors on training rollouts in one condition and tested them on separate rollouts in another. The four transfers were friction 1.2↔0.2 and Rigid↔Soft+Dissipative contact profiles. Encoders were frozen; all conditions were included in encoder training and training-data preprocessing.

We also added NSP trained only on recorded transitions, without counterfactual replays (factual-only NSP). Ordinary read-out fits across conditions; transfer fits in one and tests in another. The table averages three encoder seeds; transfer is Spearman correlation over four tests at regularization strength λ=100. Transformers received 6,000 updates.

RepresentationReservoir ordinary R²Reservoir transferTransformer transfer
ECR0.9070.9180.878
NSP0.5550.2820.099
Factual-only NSP0.8810.1280.071
CaDM-style0.7210.4370.248
LDG-style0.9400.2330.142
RMA-style0.9770.9250.925

ECR and factual-only NSP have close ordinary reservoir R², but sharply different transfer. Across six regularization strengths, reservoir ECR spans 0.74–0.92 versus factual-only NSP's 0.11–0.19. At λ=1, transformer ECR is 0.56 versus 0.49 for an untrained encoder, remaining above both NSP variants. Reservoir ECR's mean transferred numerical R² is 0.71 at λ=100 and negative at every tested λ≤10.

Longer training does not remove this distinction: at 20,000 transformer updates, ECR and factual-only NSP have ordinary R² 0.897 and 0.844, but transferred correlations 0.889 and 0.011. NSP beats ECR on all four original held-out prediction tasks. All experiments use one robot.

Q1. What does the history contain, and how were trajectories collected?

Each 79-dimensional token contains state (33), forward-speed command (1), action (12) and next-step state change (33). State contains body-frame gravity direction (3), joint positions (12), body linear/angular velocities (6) and joint velocities (12). A history contains 64 completed transitions.

One PPO locomotion policy trained across randomized slope, friction and contact conditions collected the simulated Go1 trajectories. Actions were selected deterministically while headings, commands and conditions varied. We retained 2,755 rollouts across 60 conditions, with 240 steps per rollout, an 80-step warm-up and speed commands of 0.4–1.2 m/s. We selected up to four eligible state–action queries per rollout and split by rollout.

We will add these controls, transfer results and collection details, and revise the abstract's interpretation of the original ablation.

Thank you for the additional experiments and clarifications. The normalization controls and transformer results address my concerns about training effects and the fixed reservoir. The improved transfer of steepness rankings is also interesting. However, these tests use conditions already included in encoder training, and accurate numerical estimates depend strongly on regularization. Together with the unchanged held out prediction results and inconclusive control experiment, the broader benefit remains insufficiently established for me. I therefore retain my weak rejection rating, while recognizing that the response strengthens the paper.

We thank the reviewer for the follow-up and for recognizing that the controls and transformer results address the concerns about training effects and the fixed reservoir.

The two points on transfer are accurate, and the revised manuscript states both in Section 5.6. The test uses conditions present in encoder training because it isolates the read-out: a probe fitted in one condition must still order steepness in another. The ordering result does not depend on regularization: ECR leads both NSP variants on every seed at all six tested λ (0.74–0.92 against 0.11–0.32). Only the numerical estimate depends on λ.

On held-out factor levels, ECR improves exact-cell retrieval in all ten holdouts, while NSP has lower error on the four held-out prediction tasks. We report both. We present ECR’s benefit as a transferable physical read-out learned without factor labels, not as a general prediction or control improvement.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response.

We thank the reviewers. The main new finding concerns the transfer of physical information. ECR and NSP trained only on recorded transitions have close ordinary steepness (absolute slope) read-out scores, but ECR's fitted read-out preserves steepness ordering much more reliably when friction or contact profile changes.

To address Reviewer mHqY's and Reviewer iiKy's questions about usefulness, we fitted regularized linear steepness predictors in one condition and tested them on separate rollouts in another. All conditions were available during encoder training and preprocessing. With the reservoir and regularization λ=100, mean Spearman correlation across four transfers and three encoder seeds was 0.918 for ECR versus 0.128 for NSP trained only on recorded transitions. ECR approaches the directly label-supervised RMA-style reference on this measure. The ranking advantage persists across the tested regularization strengths; numerical estimation accuracy is more sensitive.

Following the proposals from Reviewer Vxwo and Reviewer mHqY, we tested target-normalization and clipping controls. ECR's steepness read-out advantage persisted after these repairs. To address Reviewer Vxwo's architecture concern, we also repeated comparisons with a trainable transformer history encoder, using the same encoder architecture across objectives.

The main contribution is a transferable physical read-out learned from response contrasts without numerical factor-regression targets, while retaining intervention metadata, condition pairing and geometric simulator access.

We will make that contribution explicit in the abstract and introduction, add the controls and transfer results, clarify terminology, and update the AI-use statement.

AuthorsOpenReview ↗

Revised manuscript uploaded

We have uploaded the revised manuscript. Changes are coloured by reviewer: red for mHqY, green for Vxwo, purple for iiKy, and orange for changes made for all reviewers. The main changes are:

  • Sections 1 and 2 state the contribution and place ECR among label-supervised, prediction-based and intervention-based training of history representations. They also state what is new beyond the friction-only work.
  • Section 3 defines anchor, branch, donor history, receiving query and accessibility, explains the equivariance check, and states what each objective requires.
  • Section 5.2.1 and Table 1 add the target-scale controls. Section 5.5 adds the direction diagnostics.
  • Section 5.6 and Table 3 report read-out transfer across friction and contact conditions with both encoders.
  • Section 5.2 reports next-step prediction with per-seed records and the closed-loop control test.
  • Appendices J–N give the protocols and full results: target-scale controls, transformer encoder, direction diagnostics, regularization sweep, prediction and control tests, and data collection.
  • The abstract and the discussion state the revised claim and its limits. 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.