acceptodds
Under review as a conference paper at ICLR 2027

Recovering Event-Conditioned Local Symmetry Projectors for Hybrid Reinforcement Learning

Abstract

Exploiting symmetry allows reinforcement learning to share information across related states and actions. In contact-rich environments, some transformation directions remain compatible with the dynamics while others are broken. Scalar gates cannot represent these surviving subspaces, whereas fixed-basis diagonal gates discard the cross terms of rotated subspaces. We recover event-conditioned local symmetry projectors from paired reset-and-step simulator queries. Hybrid-flow residuals and typed-event information define a Gram operator whose low-defect eigenspace yields a full projector subject to finite compatibility checks. In a direct-wrench MuJoCo orbital-contact task, the projector's translation–rotation cross terms recover the orbit pivot and define a quotient shared by the actor, critics, target networks, and action transport. One projector recovered at a canonical condition per geometry was reused across eight task conditions, reducing recovery calls by 87.5% relative to condition-specific recovery. Out-of-distribution (OOD) evaluation used obstacle centers withheld from policy optimization but independently certified before training. Across 20 paired seeds, sparse recovered full improved OOD regulation AUC over no-op SAC by 0.131. It also outperformed diagonal and spectrum-matched orthogonal controls while closely matching oracle full. With matched MLPs, direct-pivot features did not reproduce the quotient benefit. After charging all recovery calls, recovered full retained a +0.037 total-call OOD AUC advantage over no-op.

Primary area:
reinforcement learning
Keywords:
reinforcement learning, local symmetry, canonicalization

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_WgZ8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes another way of exploiting proper pointwise symmetry beyond scalar gates to leverage group symmetry in reinforcement learning. They recover event-conditioned local symmetry projectors. The proposed method is novel, and the paper demonstrates promising results in the MuJoCo orbital-contact task. Overall, the paper proposes a theoretically motivated way to use proper symmetry in RL, but the validation is unclear.

Critical strengths
  • Symmetry-breaking and overcoming it in reinforcement learning is a significant problem, and the authors address the problem in a reasonable and straightforward way.
  • Through event-conditioned compatibility, they recover the local subspace and leverage the geometric information in the subspace for RL’s state-action canonicalization.
Critical weaknesses
  • Although the results were promising, the strong baseline is missing. The baselines mostly focus on ablations or standard RL approaches and do not compare with symmetry-aware RL, especially PE-SAC [1] or ApproxEquivariant RL [2], which address the same problem. Though these approaches might not support contact-rich local symmetry, the paper requires direct comparisons with previous approaches.
  • The main motivation of the paper is to properly exploit symmetry in RL. However, the benefit of a full projector in RL comes from a single MuJoCo orbital-contact task.

[1] Chang et al. "Partially equivariant reinforcement learning in symmetry-breaking environments." International Conference on Learning Representations. Vol. 2026. 2026.

[2] Park et al. "Approximate equivariance in reinforcement learning." arXiv preprint arXiv:2411.04225 (2024).

Questions for authors

Could the authors strengthen the experiments by comparing with stronger baselines and adding more RL benchmark environments to support the motivation? For example, compare PE-SAC with the candidate SE(2) algebra in the same orbital task.

3 comments
AuthorsOpenReview ↗

Response to Reviewer WgZ8

Thank you for the concrete suggestion to compare with PE-SAC.

Summary.

  • As you suggested, we ran PE-SAC with SO(2) invariant/tensor-basis networks, together with Approximate Equivariance RL and an egocentric SAC control, at matched simulator calls (Tables A–B below).
  • PE-SAC matches recovered full in total-call AUC (Δ −0.001 [−0.016, 0.013]) and success. Recovered full has higher endpoint regulation (+0.072 [0.045, 0.098]) and contact occupancy (+0.127 [0.095, 0.157]).
  • On this task, a body frame provides most of the benefit without recovery.

New experiments

All methods use seeds 9400–9419, the same splits, observations and action bounds. Results are on the 12 centers held out from policy training.

  • PE-SAC (Chang et al., 2026), built on the authors' released code, with SO(2) invariant/tensor-basis networks acting on heading, goal displacement, velocity and planar force.
  • Approximate Equivariance RL (Park et al., 2025), with a C2 half-turn action on the vector observations. This adaptation did not train well (69.0% action saturation at 50k steps).
  • Egocentric SAC: goal displacement, velocity and force are expressed in the observed body frame, with no obstacle center and no recovery queries.

Recovered full trains for 100,000 steps after 17,784 recovery calls; the other methods train for 117,784 steps. Protocols: Appendix C.3 (p. 26); learning curves: Figs. 5–6 (pp. 27–28).

Table A — Results at 117,784 simulator calls (revised Table 1, p. 8)

MethodRegulationRegulation AUCSuccessContact occupancyReturn
Recovered full0.3600.2850.4500.4953.241
Egocentric SAC0.3970.3180.7000.4454.688
PE-SAC0.2880.2870.4530.3684.841
SAC (world coordinates)0.1930.1780.2630.494−5.231
Approx. Equivariance0.1320.1290.2060.340−6.545

Means over 20 seeds and 12 centers. AUC is normalized by the total-call budget; other columns are endpoint values.

Table B — Recovered full minus comparator, 95% paired bootstrap CIs (revised Table 8, p. 27)

ComparatorOnline AUC 0–50kOnline AUC 0–100kTotal-call AUCEndpoint regulationEndpoint successEndpoint occupancy
PE-SAC−0.007 [−0.021, 0.007]+0.048 [0.031, 0.065]−0.001 [−0.016, 0.013]+0.072 [0.045, 0.098]−0.003 [−0.085, 0.075]+0.127 [0.095, 0.157]
Egocentric SAC+0.003 [−0.014, 0.019]+0.028 [0.008, 0.047]−0.033 [−0.054, −0.013]−0.037 [−0.104, 0.029]−0.250 [−0.399, −0.086]+0.050 [−0.016, 0.113]
SAC (world coordinates)+0.131 [0.119, 0.142]+0.161 [0.146, 0.177]+0.107 [0.093, 0.122]+0.167 [0.119, 0.216]+0.188 [0.076, 0.294]+0.001 [−0.051, 0.052]

The policy observation contains only free vectors and scalars, so rotation about the obstacle center acts on it exactly as rotation about any other point does, and the observed heading already gives an invariant frame. We expect the recovered projector to carry information a body frame lacks when observations include position-type quantities (absolute pose, contact or landmark positions); this remains untested.

W1 and Q1 — Stronger baselines

Compare PE-SAC with the candidate SE(2) algebra in the same orbital task.

See Tables A–B above. The policy observation contains only free vectors and scalars, and on it SO(2) captures the orbital action seen by the PE-SAC equivariant branch (Appendix C.3, p. 26). We did not run a full SE(2) variant.

W2 and Q1 — More environments

The benefit of a full projector in RL comes from a single MuJoCo orbital-contact task.

We agree, and we did not add another RL environment during the discussion period. The billiard and obstacle studies evaluate recovery only (§4.1, p. 5), and the wall study does not show an advantage for full operators (Table 13, p. 31). The revised contributions (§1) and discussion (§5) limit the claims accordingly.

I thank the authors for the detailed response. The additional results comparing with PE-SAC, Approximate Equivariance RL, and Egocentric SAC substantially address my concern regarding the lack of baselines. However, the additional results also raise new questions regarding practical benefits of the recovered projector. While the proposed approach improves over PE-SAC in terms of endpoint regulation and contact occupancy, it achieves comparable AUC and success rates. Also, Egocentric SAC outperforms the proposed approach, suggesting that the body frame representation may account for the performance gain rather than the projector. My concern regarding evaluation across diverse RL environments also remains, though the authors clarified their claims. Therefore I am leaning towards maintaining my current score.

AuthorsOpenReview ↗

Response to Reviewer WgZ8

Thank you for the follow-up. We agree that on this task the body frame accounts for the gain. Showing a practical benefit of the projector, in settings where no observed frame provides the valid symmetry and across more environments, is the main direction for our next version.

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

The paper studies partial symmetry breaking in contact-rich reinforcement learning. It proposes an event-conditioned local symmetry subspace estimator that uses paired reset-and-step simulator queries. By evaluating hybrid-flow residuals and event compatibility across candidate Lie algebra directions, the method builds a local Gram operator whose low-defect eigenspace recovers the symmetry projector. In a planar MuJoCo orbital-contact task, the authors show that the translation–rotation cross terms of the recovered rank-1 projector can decode the hidden obstacle pivot, defining a shared quotient canonicalization interface for the actor-critic networks. The paper also provides theoretical stability bounds for recovering the projector and pivot, and shows empirically that geometry-level reuse can improve out-of-distribution generalization over vanilla SAC.

Critical strengths

1.Novel connection between projector cross terms and operational coordinate canonicalization. Prior work on approximate or partially equivariant RL mostly relies on scalar or fixed-basis diagonal gating, which throws away off-diagonal cross terms. In coupled contact tasks like planar orbital contact, the translation–rotation cross terms of a full rank-1 projector actually capture the hidden pivot. Decoding this pivot to build a shared actor-critic quotient space goes beyond standard feature gating and directly cuts the effective dimensionality of the control problem.

2.Solid theoretical analysis and sample-efficient geometric reuse. The finite-query theoretical analysis neatly links directional residual fitting errors and spectral gaps to pivot estimation error via the Davis–Kahan bound. Reusing a certified projector across different starting phases and travel directions also significantly cuts down simulator queries, offering a clear way to amortize probing costs.

Critical weaknesses

1.Lack of external baselines in empirical evaluation. Although the paper discusses several recent partially equivariant RL methods, symmetry discovery techniques, and learned canonicalization frameworks in related work, none are tested in the MuJoCo experiments. All comparisons are against internal ablations or degenerate variants of the proposed method. Without benchmarking against existing gate-based equivariant RL or data-driven canonicalization baselines, it is hard to tell whether the probing framework really brings a practical advantage over existing alternatives.

2.Heavy assumptions regarding simulator access, oracle geometry IDs, and idealized shapes. First, the estimator relies on active reset-and-step query access to a deterministic simulator. Second, during online RL, the agent does not infer the projector directly from observations, but instead looks up pre-certified projectors using exact ground-truth geometry IDs. Relying on ground-truth IDs at test time weakens the claims around out-of-distribution generalization. Third, decoding a static pivot from the projector cross terms assumes a circular obstacle with constant curvature. It remains unclear whether this formulation extends to non-circular or irregular shapes where the instantaneous center of rotation drifts along the contact surface.

3.Dense writing and heavy jargon. The text is very condensed, with many shorthand phrases used as nouns without being clearly introduced or defined first.

Questions for authors

1.Could you provide empirical comparisons against multiple existing baselines, including partially equivariant RL approaches and learned canonicalization methods, on the orbital-contact task?

2.The current online setup uses exact geometry identifiers to look up pre-certified projectors. If the environment does not provide these ground-truth identifiers, how would the agent select or interpolate the correct projector during OOD evaluation?

3.How would the pivot decoding and geometric reuse handle non-circular obstacles where the instantaneous center of rotation continuously drifts along the contact surface?

4 comments
AuthorsOpenReview ↗

Response to Reviewer btBR (1/2)

Thank you for the specific questions on baselines, identifiers and obstacle shape.

Summary.

  • We added PE-SAC, Approximate Equivariance RL and an egocentric SAC control at matched simulator calls (Tables A–B below).
  • PE-SAC matches recovered full in total-call AUC and success; recovered full has higher endpoint regulation (+0.072 [0.045, 0.098]).
  • Egocentric SAC, using neither the obstacle center nor recovery, exceeds recovered full once recovery calls are charged (total-call AUC +0.033, success +0.25).
  • On this task, recovery is validated but not shown to be needed for control. The abstract, contributions (§1) and discussion (§5) are revised accordingly.

New experiments

All methods use seeds 9400–9419, the same splits, observations and action bounds. Results are on the 12 centers held out from policy training.

  • PE-SAC (Chang et al., 2026), built on the authors' released code, with SO(2) invariant/tensor-basis networks acting on heading, goal displacement, velocity and planar force.
  • Approximate Equivariance RL (Park et al., 2025), with a C2 half-turn action on the vector observations. This adaptation did not train well (69.0% action saturation at 50k steps).
  • Egocentric SAC: goal displacement, velocity and force are expressed in the observed body frame, with no obstacle center and no recovery queries.

Recovered full trains for 100,000 steps after 17,784 recovery calls; the other methods train for 117,784 steps. Protocols: Appendix C.3 (p. 26); learning curves: Figs. 5–6 (pp. 27–28).

Table A — Results at 117,784 simulator calls (revised Table 1, p. 8)

MethodRegulationRegulation AUCSuccessContact occupancyReturn
Recovered full0.3600.2850.4500.4953.241
Egocentric SAC0.3970.3180.7000.4454.688
PE-SAC0.2880.2870.4530.3684.841
SAC (world coordinates)0.1930.1780.2630.494−5.231
Approx. Equivariance0.1320.1290.2060.340−6.545

Means over 20 seeds and 12 centers. AUC is normalized by the total-call budget; other columns are endpoint values.

Table B — Recovered full minus comparator, 95% paired bootstrap CIs (revised Table 8, p. 27)

ComparatorOnline AUC 0–50kOnline AUC 0–100kTotal-call AUCEndpoint regulationEndpoint successEndpoint occupancy
PE-SAC−0.007 [−0.021, 0.007]+0.048 [0.031, 0.065]−0.001 [−0.016, 0.013]+0.072 [0.045, 0.098]−0.003 [−0.085, 0.075]+0.127 [0.095, 0.157]
Egocentric SAC+0.003 [−0.014, 0.019]+0.028 [0.008, 0.047]−0.033 [−0.054, −0.013]−0.037 [−0.104, 0.029]−0.250 [−0.399, −0.086]+0.050 [−0.016, 0.113]
SAC (world coordinates)+0.131 [0.119, 0.142]+0.161 [0.146, 0.177]+0.107 [0.093, 0.122]+0.167 [0.119, 0.216]+0.188 [0.076, 0.294]+0.001 [−0.051, 0.052]

The policy observation contains only free vectors and scalars, so rotation about the obstacle center acts on it exactly as rotation about any other point does, and the observed heading already gives an invariant frame. We expect the recovered projector to carry information a body frame lacks when observations include position-type quantities (absolute pose, contact or landmark positions); this remains untested.

AuthorsOpenReview ↗

Response to Reviewer btBR (2/2)

W1 and Q1 — External baselines

Could you provide empirical comparisons against multiple existing baselines, including partially equivariant RL approaches and learned canonicalization methods, on the orbital-contact task?

See Tables A–B above. We did not add a learned canonicalization method. Egocentric SAC is an analytic canonicalization, and it performs at least as well as recovered full once recovery cost is included.

W2 and Q2 — Geometry identifiers

If the environment does not provide these ground-truth identifiers, how would the agent select or interpolate the correct projector during OOD evaluation?

You are right that online lookup uses exact geometry keys (Algorithm 1, p. 17) and that all evaluation centers were queried before training. Without identifiers, recovery would need to run at the current geometry before the quotient applies, at 741 calls per geometry (Table 14, p. 32). We have not tested this.

Predicting the projector directly from observations is harder: in the wall study, a network trained on recovered labels reaches a 32.6° median OOD angle (Table 12, p. 31). With the 20 frozen 50k-step policies, withholding half of the 12 held-out keys reduces regulation from 0.472 to 0.358 (paired change −0.114 [−0.137, −0.089]). The full coverage sweep is in Appendix C.4 (Table 9, p. 28).

W2 and Q3 — Non-circular obstacles

How would the pivot decoding and geometric reuse handle non-circular obstacles where the instantaneous center of rotation continuously drifts along the contact surface?

A fixed-center rotation preserves the dynamics only for a rotationally symmetric contact surface, so the pivot table and its reuse across phases depend on a circular obstacle. A drifting center would require context-dependent recovery and a control interface in place of the per-geometry pivot table. We have not tested this.

W2 — Simulator access

The method assumes deterministic reset-and-step access and typed event descriptors (§3.1; Appendices A.1–A.3).

W3 — Writing

The revision defines whitening and the rotational anchor at first use (§3.1, p. 3; §3.4, p. 4), defines certification as acceptance under the finite checks (Figure 1 caption, p. 3), and expands Figure 1c to explain the two control interfaces. Algorithm 1 (p. 17) gives the complete procedure.

I thank the authors for providing the supplementary experiments and detailed explanations. I greatly appreciate the authors’ openness and candor in defining the scope and limitations of the method. These explanations and the supplementary experiments have helped me assess its performance more thoroughly; however, the method still lacks sufficient data support for handling cases such as non-circular objects and OOD scenarios. I suggest that, in future revisions, the authors adopt more intuitive data visualizations to illustrate the core mechanisms and characteristics of the method, which would greatly help readers understand its advantages and practical applicability boundaries. Based on the current experimental results, I have decided to maintain my score of 2.

AuthorsOpenReview ↗

Response to Reviewer btBR

Thank you for the constructive suggestion. We agree that the current results do not cover non-circular contact surfaces or geometries never queried before training, and the revision states both as limitations. We will add visualizations of the recovered projector, the decoded pivot and the resulting control frame across contact configurations, so that readers can see the mechanism and where it applies.

Reviewer_6beUOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
1
Soundness
1
Read review
Paper summary

The paper estimates locally preserved transformation directions within a supplied family of transformations, using paired simulator queries and contact-event checks. In the principal circular-obstacle task, the resulting projection matrix determines a rotation center. Observations and actions are then expressed relative to radial and tangential axes around that center. The paper provides numerical error analysis, recovery experiments, and reinforcement-learning evaluations.

Critical strengths
  1. The numerical formulation and analysis: The numerical error bounds appear sound under their stated assumptions. They show that sufficiently accurate numerical measurements produce sufficiently accurate subspaces and centers provided the relevant smoothness, conditioning, and eigenvalue separation assumptions hold.
  2. Experimental controls: The principal learning experiment uses multiple seeds, an oracle analytic-projector reference, controlled matrix ablations, and a comparison that charges the simulator calls used for recovery.
Critical weaknesses
  1. The principal experiment does not directly test the comparison with gating methods that motivates the paper. The orbital diagonal control removes matrix entries before extracting the center, producing an origin-centered coordinate system (Appendix A.6). This tests the chosen center-extraction procedure, not learned gating methods. The wall study does include scalar and diagonal networks, but they are trained with oracle or recovered projector labels and frozen during policy training. Moreover, Table 10 reports the highest mean learning-curve area for recovered-label diagonal gates. The results do not show a consistent practical advantage for full matrices.

  2. The need for recovery and its usefulness beyond this benchmark remain unclear. The obstacle center determines the projector analytically, as the analytic reference demonstrates. This makes the experiment useful for numerical validation, but leaves the advantage over using available geometry unresolved. The feature comparison supports rotating observations and actions rather than merely supplying center-relative position; it does not establish the necessity of the estimation procedure. Extracting a center from a one-dimensional rotational subspace is also specialized. Other transformation families or dimensions would require a different construction for using the estimate in control. The paper should distinguish the general numerical estimation procedure from the additional task-specific construction needed to make its output useful for control.

  3. The acceptance criteria need better justification. How should the reference length, perturbation sizes, eigenvalue cutoff, angular tolerance, reward tolerance, and penetration limit be chosen on another task? The connection between the finite checks and the assumptions behind the numerical bounds needs clarification.

  4. The generalization experiment has limited scope. Every evaluation center is queried before policy training, and exact geometry identifiers select stored matrices. The results therefore do not demonstrate generalization to unseen obstacle locations.

  5. The connection between the framing and the technical contribution needs clarification. The title and framing suggest a contribution to hybrid reinforcement learning, while the central technical development concerns numerical symmetry estimation: recovering transformation directions from simulator queries and bounding estimation errors. The control experiments demonstrate an application of these estimates through transformed observations and actions, or critic regularization. What general advance in reinforcement learning follows beyond these applications?

  6. The writing is very difficult to follow. Terms and symbols are often insufficiently defined at first use, with essential explanations deferred to the appendix. “Whitening,” “rotational anchor,” and “certification” also need clearer explanations. Figure 1c also combines two different uses of the estimated matrix—coordinate transformation and critic regularization—which deserve a more concrete explanation. The existing main example should be explained as one continuous procedure, with definitions provided when needed and more standard terminology throughout.

Questions for authors
  1. Can you compare against learned scalar and fixed-basis diagonal gating methods on the principal task, using appropriate implementations and matched training and simulator-call budgets?
  2. Can you demonstrate control benefits in a setting where the desired transformation is not immediately available from a known obstacle center? How would the estimated subspace be used when its dimension or geometric interpretation differs from the orbital case?
  3. How were the reference length and acceptance tolerances selected? Which assumptions of the numerical analysis do the finite checks establish, and which remain assumptions?
  4. How does control performance change as acceptance coverage decreases or previously unqueried "geometries" are introduced?
  5. What motivates the particular composite score combining contact occupancy, angular progress, and final goal accuracy?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 6beU (1/4)

Thank you for the careful reading, including the wall results and Appendix A.6.

Summary.

  • We added PE-SAC, Approximate Equivariance RL and an egocentric SAC control at matched simulator calls (Tables A–B below).
  • PE-SAC matches recovered full in total-call AUC (Δ −0.001 [−0.016, 0.013]) and success. Recovered full has higher endpoint regulation (+0.072 [0.045, 0.098]).
  • Egocentric SAC uses no obstacle center or recovery and exceeds recovered full once recovery calls are charged: total-call AUC +0.033 [0.013, 0.054], success +0.25 [0.086, 0.399].
  • We agree with your central point: on the orbital task, the experiments validate recovery and decoding but do not show that recovery is needed for control. The abstract, contributions (§1) and discussion (§5) are revised accordingly.

New experiments

All methods use seeds 9400–9419, the same splits, observations and action bounds. Results are on the 12 centers held out from policy training.

  • PE-SAC (Chang et al., 2026), built on the authors' released code, with SO(2) invariant/tensor-basis networks acting on heading, goal displacement, velocity and planar force.
  • Approximate Equivariance RL (Park et al., 2025), with a C2 half-turn action on the vector observations. This adaptation did not train well (69.0% action saturation at 50k steps).
  • Egocentric SAC: goal displacement, velocity and force are expressed in the observed body frame, with no obstacle center and no recovery queries.

Recovered full trains for 100,000 steps after 17,784 recovery calls; the other methods train for 117,784 steps. Protocols: Appendix C.3 (p. 26); learning curves: Figs. 5–6 (pp. 27–28).

Table A — Results at 117,784 simulator calls (revised Table 1, p. 8)

MethodRegulationRegulation AUCSuccessContact occupancyReturn
Recovered full0.3600.2850.4500.4953.241
Egocentric SAC0.3970.3180.7000.4454.688
PE-SAC0.2880.2870.4530.3684.841
SAC (world coordinates)0.1930.1780.2630.494−5.231
Approx. Equivariance0.1320.1290.2060.340−6.545

Means over 20 seeds and 12 centers. AUC is normalized by the total-call budget; other columns are endpoint values.

Table B — Recovered full minus comparator, 95% paired bootstrap CIs (revised Table 8, p. 27)

ComparatorOnline AUC 0–50kOnline AUC 0–100kTotal-call AUCEndpoint regulationEndpoint successEndpoint occupancy
PE-SAC−0.007 [−0.021, 0.007]+0.048 [0.031, 0.065]−0.001 [−0.016, 0.013]+0.072 [0.045, 0.098]−0.003 [−0.085, 0.075]+0.127 [0.095, 0.157]
Egocentric SAC+0.003 [−0.014, 0.019]+0.028 [0.008, 0.047]−0.033 [−0.054, −0.013]−0.037 [−0.104, 0.029]−0.250 [−0.399, −0.086]+0.050 [−0.016, 0.113]
SAC (world coordinates)+0.131 [0.119, 0.142]+0.161 [0.146, 0.177]+0.107 [0.093, 0.122]+0.167 [0.119, 0.216]+0.188 [0.076, 0.294]+0.001 [−0.051, 0.052]
AuthorsOpenReview ↗

Response to Reviewer 6beU (2/4)

W2 and Q2 — Need for recovery and use beyond the orbital case

The need for recovery and its usefulness beyond this benchmark remain unclear.

How would the estimated subspace be used when its dimension or geometric interpretation differs from the orbital case?

We agree. The policy observation contains only free vectors (heading, goal displacement, velocity) and scalars, so rotation about the obstacle center acts on it exactly as rotation about any other point does, and the observed heading already gives an invariant frame. The paper therefore contains no setting where recovery is shown to be necessary for control.

We expect recovery to matter when observations include position-type quantities, where the valid symmetry is a proper subspace of se(2) that depends on unobserved geometry. We have not tested this.

The estimator outputs a projector field. Section 3.4 (p. 5) now separates it from the task-specific control interface, with orbital pivot decoding and wall critic regularization as the two examples.

W1 and Q1 — Comparison with gating methods

The principal experiment does not directly test the comparison with gating methods that motivates the paper.

PE-SAC is now our learned gated comparator at matched simulator calls (Tables A–B above). We did not add a standalone learned fixed-basis diagonal gate on the orbital task.

The orbital diagonal control supplies the exact diag(P̂), the best output any fixed-basis diagonal gate could produce in this interface. It shows that the pivot cannot be decoded without the cross terms, but it does not evaluate gating methods in their native use, which PE-SAC now covers.

On the wall task, you are right on both points: recovered-label diagonal gates have the highest OOD AUC (0.620), and the geometry oracle's final OOD success (0.524) is below vanilla SAC (0.547). These results do not support a practical advantage for full operators (Appendix C.7, Table 13, p. 31).

W4 and Q4 — Coverage and unqueried geometries

Every evaluation center is queried before policy training, and exact geometry identifiers select stored matrices.

Correct. The results show that the policy generalizes to centers held out from its training, not that recovery generalizes to new geometries. Here, OOD denotes centers held out from policy training (Appendix C.2, pp. 22–23).

We evaluated reduced table coverage with the 20 frozen 50k-step policies on the 12 held-out centers. Withheld keys use the identity fallback (Appendix C.4, Table 9, p. 28):

Missing keysRegulationSuccessContact occupancyReturn
0%0.4720.5470.5642.041
25%0.4150.4730.5970.720
50%0.3580.3990.629−0.601
75%0.3010.3240.662−1.922
100%0.2440.2500.695−3.243

Withholding all keys changes regulation by −0.228 [−0.274, −0.178]. Contact occupancy rises while success and return fall, so maintaining contact alone does not complete the task. Intermediate rows average over all key subsets, so they interpolate linearly.

At a geometry never queried, recovery must run first, at 741 simulator calls per geometry (Table 14, p. 32). We have not tested online recovery at new geometries.

AuthorsOpenReview ↗

Response to Reviewer 6beU (3/4)

W3 and Q3 — Acceptance criteria and the theory

How were the reference length and acceptance tolerances selected? Which assumptions of the numerical analysis do the finite checks establish, and which remain assumptions?

Appendix C.2 (p. 23) gives the rationale for each setting:

  • Reference length L = 0.35 m: the radius of the obstacle-center layout, putting translation and rotation on a common length scale in M=diag⁡(1,1,L2)M=\operatorname{diag}(1,1,L^2). With the same L in whitening and decoding, the pivot decoded from an exact orbital projector is independent of L.
  • Probe radii {0.0004, 0.0008} m: a factor-of-two comparison of the local response. In the 48-context audit, each radius alone recovered the correct rank in every context (P90 errors 1.13° and 0.91°).
  • Relative spectral threshold 0.015: all thresholds from 0.0025 to 0.02 select the correct rank in all 48 audit contexts (Fig. 7a, p. 29).
  • Radius-agreement tolerance 10°: allows finite-scale variation between the two radius projectors. The maximum observed disagreement was 0.99° over the 24 sparse keys and 1.68° over the 192 dense conditions.
  • Reward tolerance τ_r = 10⁻⁴: an absolute budget, in reward units, on fitted first-order reward variation along the recovered subspace: ϵmax⁡∥Pˉgˉ^r∥2≤10−4\epsilon_{\max}\|\bar P\widehat{\bar g}_r\|_2\leq10^{-4}.
  • Penetration limit 0.01 m: the benchmark's existing mechanics limit.
  • Robustness: with 25 instead of 37 directions, or up to 10% residual noise, rank accuracy stays at 100% (Appendix C.5, p. 29).

For another task: set L from the task's geometric scale; choose probe radii that preserve the event regime while resolving differences above simulator error; set spectral and angular tolerances from cross-radius stability; set reward and penetration tolerances from the permitted reward variation and physical overlap (Appendix C.2, p. 23).

What each check establishes:

  • Design-rank screening tests the full-row-rank condition of Proposition 1.
  • Event-signature screening assesses the shared event stratum of Assumption A.1.
  • Rank and angle agreement across radii probes finite-difference sensitivity under Assumption A.2.
  • Smoothness, numerical error bounds and the population eigengap remain assumptions.
  • The reward check bounds fitted first-order reward variation along the recovered subspace (Eq. 8, p. 4).
  • Assumption A.3 (local sampling) applies to the obstacle and wall estimators; the orbital estimator uses Proposition 1 with a pooled design.

Section 3.4 (p. 5) now states these connections. Assumptions A.1–A.3 are in Appendix A.7.2 (p. 18).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.