acceptodds
Under review as a conference paper at ICLR 2027

Equivariance Begins at the Source: Characterizing Compatible Sources for Generative Robot Policies

Abstract

Equivariant transport alone does not guarantee an equivariant manipulation policy. A compatible source distribution–one that transforms consistently with the observation–is required. We show that, under an invertible equivariant flow, the policy inherits the source's equivariance error exactly and therefore cannot correct an incompatible source. In particular, the commonly used isotropic Gaussian centered at the world-frame origin is not translation-equivariant—the KL divergence between the source and its translate grows quadratically with translation distance. We further characterize all compatible conditional sources through a unique frame-relative normal form. For robot actions, this characterization shows that translation requires observation dependence, whereas normalized Haar measure is the unique observation-independent rotational source. Building on these results, we propose GARLIC, an -equivariant flow-matching policy that anchors its translational source at the end-effector, samples rotations from the normalized Haar measure, and uses an equivariant geometric-algebra transformer for transport. On RLBench, GARLIC improves task success and out-of-distribution generalization over existing baselines, demonstrates greater parameter efficiency, and transfers to real-robot manipulation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Equivariance, Flow-matching policy, robotic manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_HnKMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies source distributions in equivariant generative robot policies and argues that an equivariant transport mechanism alone is insufficient to guarantee an equivariant policy. The paper formalizes source compatibility, shows that equivariance error in the source is preserved by an invertible equivariant flow, and characterizes compatible source distributions for translation, rotation, and gripper components. Based on this analysis, the authors introduce GARLIC, which uses an end-effector-anchored translational source, an observation-independent rotational source corresponding to the Haar measure on SO(3), and an equivariant geometric-algebra flow-matching policy.

However, I have several substantial concerns regarding the practical novelty and the empirical interpretation of the method. In particular, the current experiments do not sufficiently isolate whether the observed gains come from source compatibility itself or simply from using a more informative and easier conditional action prior whose mass is located closer to the demonstrated target actions. In addition, the practical benefit over standard coordinate canonicalization / relative-action formulations is not yet clearly established, and the real-world and OOD evaluations are too limited to support some of the broader generalization claims.

Critical strengths

Please identify one or two critical strengths of the paper that support acceptance. For each, explain briefly why it is consequential to your overall assessment. The goal is not to enumerate every positive point, but to identify and substantiate the considerations that most strongly determine your recommendation. Enter “None” if no substantive strength can be identified.

The paper identifies an important and often overlooked component of equivariant generative policies. The observation that end-to-end equivariance depends not only on the denoiser / velocity field but also on the source distribution is conceptually useful. The separation between source compatibility and equivariant transport provides a clear framework for analyzing generative robot policies.

Critical weaknesses

Please identify one or two critical weaknesses of the paper that support rejection. For each, explain briefly why it is consequential to your overall assessment. The goal is not to enumerate every negative point, but to identify and substantiate the considerations that most strongly determine your recommendation. Enter “None” if no material weakness can be identified.

1. The practical advantage over coordinate canonicalization or relative-action formulations is unclear

The proposed translational source is anchored at the current end-effector position. However, in robotic manipulation, world-frame and end-effector-frame actions can be converted through the current robot pose. Therefore, the proposed source construction appears closely related to defining the action generation process in an end-effector-relative coordinate system and then transforming the result back to the world frame.

The paper provides a useful theoretical interpretation of this design, but it is unclear whether GARLIC offers a practical advantage over conventional canonicalization or relative-action representations.

A matched comparison between:

  • GARLIC with a world-frame anchored source, and
  • an end-effector-frame canonicalized policy using the same or similar backbone

would be important for establishing the practical novelty of the method.

2. Source compatibility is confounded with a better and easier action prior

GARLIC centers its translational source at the current end-effector position instead of the world origin. This change does two things simultaneously:

  • it improves source compatibility with geometric transformations;
  • it places the source closer to typical robot actions.

The second factor can independently make the flow-matching problem easier. A source initialized near the current end-effector may require a shorter and simpler transport to reach the demonstrated action trajectory. This can improve optimization, sample efficiency, and finite-step inference even without considering equivariance.

Therefore, the comparison between a fixed world-origin Gaussian and an end-effector-anchored source does not isolate the effect of source compatibility. The paper would be significantly stronger if it included controlled experiments using multiple compatible sources with different distances to the target actions, or compatible and incompatible sources with approximately matched source-target distances. Without such experiments, it remains unclear whether the performance gains mainly come from equivariance or from using a more favorable conditional action prior.

Figure 1 illustrates an object-anchored source, while GARLIC actually anchors the source at the robot end-effector. An object-anchored source can already provide task-relevant spatial information because the source is located near the manipulated object. This makes it difficult to separate the benefit of equivariance from the benefit of placing the source near the desired action region.

Questions for authors

Ask only questions whose answers could materially affect your assessment. Distinguish clarification requests from suggestions for future work.

#1. How much of GARLIC's improvement comes from source compatibility, and how much comes from placing the source closer to the target action distribution?

#2. Can the authors compare several compatible sources that have substantially different distances to the demonstrated actions?

#3. Can the authors compare GARLIC against an end-effector-frame canonicalized policy using a matched architecture?

4 comments

Thanks for your careful reading and the insightful suggestions. We answer each question below. All success rates are computed over 15 checkpoints: the last 5 epochs of each of 3 training seeds.

Q1. Contributions of source compatibility and proximity

We agree that the contributions of compatibility and proximity should be better isolated. We ran both controlled experiments proposed in the review: compatible and incompatible sources at matched distances (this section), and compatible sources at different distances (Q2). Here we compare GARLIC with 3 alternative source distributions:

  • Shifted-anchor: N(pee+0.5Reeez,σ2I)\mathcal{N}(p_{ee}+0.5R_{ee}e_z,\sigma^2I), a compatible source whose center is shifted 0.5 m from the end-effector along its local z-axis.
  • Mean-centered: N(pˉ,σ2I)\mathcal{N}(\bar p,\sigma^2I), an observation-independent Gaussian centered at the mean training action pˉ\bar p.
  • Fixed Gaussian: N(0,σ2I)\mathcal{N}(0,\sigma^2I), the observation-independent source of Table 4.

We measure proximity as the mean distance between each source center and the corresponding demonstrated action position. The table reports in-distribution (ID) and out-of-distribution (OOD) success rates, with Δ=OOD−ID\Delta=\mathrm{OOD}-\mathrm{ID}.

SourceCompatibilityDistance (m)IDOODΔ\Delta
GARLIC✓0.23084.7±5.183.2±4.7−1.5
Shifted-anchor✓0.42982.4±1.482.7±2.8+0.3
Mean-centered✗0.20574.2±2.763.4±6.9−10.8
Fixed Gaussian✗0.47550.7±5.531.3±5.0−19.4

In short, at matched distance compatibility accounts for 10.5 to 31.7 points in ID and 19.8 to 51.4 in OOD, whereas proximity accounts for 2.3 points when the source is compatible and 23.5 when it is not.

Source compatibility contributes to performance beyond proximity to the target actions. At comparable distances, the compatible sources substantially outperform their incompatible counterparts. Shifted-anchor outperforms Fixed Gaussian by 31.7 points in ID and 51.4 in OOD, despite their similar distances (0.429 m vs 0.475 m). Similarly, GARLIC outperforms Mean-centered by 10.5 points in ID and 19.8 in OOD, despite having a larger mean distance to the targets (0.230 m vs 0.205 m). The compatible sources also exhibit smaller ID–OOD performance gaps, consistent with the benefits of equivariance.

The effect of proximity depends on compatibility. Between the two compatible sources, increasing the distance from 0.230 m (GARLIC) to 0.429 m (Shifted-anchor) reduces ID success by only 2.3 points. In contrast, moving the incompatible Fixed Gaussian from the origin to the mean training action reduces its distance from 0.475 m to 0.205 m and improves ID success from 50.7% to 74.2%. This supports the reviewer's point that a source closer to the targets can make the flow-matching problem easier, consistent with prior work on informative sources (Lee et al., ICLR 2022; Chen et al., RSS 2024). Proximity therefore matters mainly when compatibility is absent.

We attribute this pattern to sample efficiency. A compatible source makes the policy equivariant, which improves sample efficiency (Wang et al., CoRL 2024; Yang et al., CoRL 2024), so the policy learns well wherever the source is placed. Without compatibility this benefit is lost, and an informative source helps instead.

We added this comparison as Section 6.6, with details in Appendix D.4.

Q2. Compatible sources at different distances

We compare 3 compatible sources at different distances from the demonstrated actions:

  • Centroid-anchor: N(c(X),σ2I)\mathcal{N}(c(X),\sigma^2I), at the centroid c(X)c(X) of the point cloud.
  • GARLIC: N(pee,σ2I)\mathcal{N}(p_{ee},\sigma^2I), at the end-effector.
  • Shifted-anchor: N(pee+0.5Reeez,σ2I)\mathcal{N}(p_{ee}+0.5R_{ee}e_z,\sigma^2I), 0.5 m from the end-effector along its z-axis.
SourceDistance (m)IDOODΔ\Delta
Centroid-anchor0.18084.9±3.083.5±2.1−1.4
GARLIC0.23084.7±5.183.2±4.7−1.5
Shifted-anchor0.42982.4±1.482.7±2.8+0.3

The 3 compatible sources perform comparably despite very different distances to the targets. Their mean distances range from 0.180 m to 0.429 m, while ID success varies by at most 2.5 points and OOD success by less than 1 point. In particular, Centroid-anchor is closer to the targets than GARLIC but performs the same. Among compatible sources, proximity thus has little effect on performance. This is consistent with the explanation in Q1.

Centroid-anchor also clarifies Figure 1. The object-anchored source in the figure illustrates the general construction of Corollary 2, which permits any equivariant anchor, including an object position, the point-cloud centroid, or the end-effector position. GARLIC uses the end-effector as a readily available anchor, while Centroid-anchor incorporates scene geometry directly into the source center. Their comparable performance shows that the benefit of a compatible source does not depend on anchoring at the end-effector or on placing the source close to the targets.

Q3. Comparison with end-effector-frame canonicalization

Thank you for suggesting this comparison. We evaluate coordinate-frame canonicalization using 2 backbones, DP3 and GATr. To separate the roles of the frame’s origin and its orientation, we fix the origin at the end-effector position peep_{ee}, while its orientation is defined by the world axes, end-effector axes ReeR_{ee}, or a random rotation sampled uniformly from SO(3)SO(3) for every training sample and policy call. The table (Table 8 in the revision) reports ID/OOD success rates (%).

BackboneWorld axesEnd-effector axes ReeR_{ee}Random axes
DP344.3±6.7 / 9.3±6.442.7±1.2 / 41.3±3.039.6±8.5 / 41.3±7.0
GATr84.7±5.1 / 83.2±4.783.7±3.9 / 84.2±4.584.3±4.3 / 84.6±4.8

With the equivariant GATr backbone, canonicalization and GARLIC achieve comparable performance. The world-axes entry is GARLIC itself: a frame at peep_{ee} with no rotation places the source at N(pee,σ2I)\mathcal N(p_{ee},\sigma^2 I), which is the anchored source. The end-effector-axes entry is the canonicalization suggested in the review. We agree with the reviewer that the two are closely related. With an equivariant backbone they are in fact the same policy: the frame orientation cancels and only its origin remains (Remark 2, Appendix A.6 of the revision).

To see this, consider a frame with origin peep_{ee} and orientation Q∈SO(3)Q\in SO(3). Sampling the translational source in this frame as p~0∼N(0,σ2I)\tilde p_0\sim\mathcal N(0,\sigma^2I) and transforming it to world coordinates gives

\sim\mathcal N(p_{ee},\sigma^2I).$$ Because the Gaussian is isotropic, this distribution does not depend on $Q$. Similarly, the 6D rotational source $\mathcal N(0,\sigma^2 I_6)$ is unchanged by $Q$, so the induced rotation remains Haar-distributed. Thus, every frame orientation induces the same source, which is GARLIC's anchored source. Since GATr is equivariant, canonicalization (expressing the observation in the frame, generating there, and mapping the action back) gives the same policy as generating directly in the world frame from this source. In other words, with an equivariant backbone, the anchored source is a canonicalization that needs only an anchor point and no orientation. The table agrees: the three GATr entries are separately trained models, and their success rates match across frame orientations. **With the non-equivariant DP3 backbone, the frame orientation matters.** Without an equivariant backbone, the orientation no longer cancels. With world axes, OOD success falls to 9.3%. Making the orientation equivariant, either exactly with the end-effector axes or in distribution with a random rotation, removes the ID–OOD gap (41.3% OOD in both cases), but success remains about half that of GATr. **Practical implications.** With a matched equivariant backbone, GARLIC and end-effector-frame canonicalization perform the same; the practical difference is in what each requires. End-effector-frame canonicalization needs a full pose to define its frame, whereas the anchored source needs only a single equivariant point. Centroid-anchor (Q2) shows this directly: the point-cloud centroid defines no orientation, yet it matches GARLIC (84.9 / 83.5 vs 84.7 / 83.2). This matters in settings where no end-effector pose is observed, such as generating grasp poses from an object point cloud: the centroid still provides an anchor, and no frame orientation has to be estimated. The source view thus shows what canonicalization actually requires: it separates the frame into an origin and an orientation, and identifies when each is needed. That an anchor alone can suffice is implicit in practices such as centroid centering (Yang et al., CoRL 2024); our analysis states it explicitly and gives the condition under which it holds. The reviewer's question led us to make this explicit; the comparison is now in Section 6.7 (Table 8) of the revision, with the formal statement in Remark 2 and its proof in Appendix A.6. The review also mentions relative-action formulations. We ran this comparison with the same GATr backbone (65.5 / 64.7 versus 84.7 / 83.2 for GARLIC); details are in our response to Reviewer eGSi (W1/Q1) and in Section 6.5 (Table 6) of the revision. ## Scope of the OOD and real-world evaluations The OOD experiments test robustness to $SE(2)$ transformations of the whole scene. The real-robot experiments test generalization to unseen object placements with a fixed base, and deployment on physical hardware. We revised the abstract and the conclusion to state this scope explicitly. Beyond the original OOD results on four tasks, the experiments added in this response use the same OOD evaluation to examine the importance of source compatibility (Q1), the difference between a compatible source and an informative one (Q1, Q2), and what canonicalization requires of the frame orientation (Q3). Together they provide further empirical evidence for our theory.
Reviewer_HnKMOpenReview ↗

Response To Authors

The author addressed some of my concerns, and I am inclined to raise our rating.

Thank you for your time and thoughtful review. The revised manuscript includes the experiments and discussions you suggested, which helped us improve our paper.

Reviewer_eGSiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how the source distribution affects the equivariance of flow-matching robot policies. The authors show that, under an equivariant and invertible flow, the policy inherits the source distribution’s equivariance error exactly. In particular, the commonly used Gaussian source centered at the world origin violates translation equivariance when actions are represented as absolute world-frame poses. They further characterize the conditions for source compatibility, demonstrating that translation equivariance requires an observation-dependent source, whereas the Haar measure suffices for rotation equivariance. Based on these findings, they propose GARLIC, which anchors the translational source at the end-effector and employs a GATr backbone. Experiments across four RLBench tasks and four real-world robotic tasks demonstrate improved success rates and greater robustness to SE(2) scene shifts.

Critical strengths

I find the motivation of this paper valuable. It addresses a question that has received relatively little attention in equivariant robot policies: how the choice of source distribution affects equivariance when actions are represented as absolute poses in the world frame. Making the source an explicit part of the equivariance analysis is a useful perspective.

The paper is also well written and thorough. Despite the number of theorems, propositions, and lemmas, the arguments are logically consistent and clearly connected to the proposed method. The controlled planar experiment provides convincing evidence supporting the theoretical claims, and the discussion of how the results relate to canonicalization and existing policies adds useful context.

Critical weaknesses

My main concern is that the proposed source design is equivalent to relative translation and the paper lacks a controlled baseline to isolate its contribution. The source analysis is valuable, but its practical implication is limited. With world-frame relative translations (e.g., Δp=pa−pee)\Delta p = p_a − p_{ee}), a fixed isotropic Gaussian source is already compatible, since Δp\Delta p is invariant to translation and only rotates with the scene. GARLIC’s anchored source N(pee,σ2I)\mathcal{N}(p_{\mathrm{ee}}, \sigma^2 I) is mathematically equivalent to this reparameterization, so the same effect can be obtained simply through a common action representation. Although EquiBot and DP3-Canon handle translation through velocity or canonicalized actions, they also use different backbones, so it remains unclear whether GARLIC’s gains come from the source design or from GATr. A controlled baseline with the same GATr backbone, relative translation actions, and a fixed Gaussian source is missing. The paper also does not justify why the absolute world-frame formulation is preferable.

Questions for authors
  1. Could the authors compare against the same GATr backbone with relative translation actions in the world frame (e.g., w.r.t. the current end-effector) and a fixed Gaussian source? This would clarify whether the anchored source offers any benefit beyond this common reparameterization, and why the absolute world-frame formulation is preferable.

  2. In Table 4, the fixed Gaussian reduces ID success from 84.7% to 50.7%, a larger drop than its ID–OOD gap (−19.4). Since Theorem 1 concerns equivariance error rather than in-distribution fit, part of the gain may instead come from the anchored source being closer to the target actions, which is also suggested by the non-equivariant PFM backbone improving from 54.2% to 61.7% ID with the anchored source. Could the authors disentangle the two effects, e.g., with an observation-independent Gaussian centered at the mean training action?

  3. In the OOD evaluation, which parts of the observation are transformed: only the task objects, or also the initial end-effector pose? If the end-effector is not transformed, the anchor does not move with the objects, and the theoretical guarantee does not directly apply. Even if it is, the rest of the robot in the point cloud remains fixed, so the observation is still not exactly transformed by the group action. Could the authors clarify the setup and discuss how this relates to real-world deployment, where the robot base is fixed and the initial end-effector pose cannot be transformed together with the scene?

3 comments

Thank you for the positive feedback and the constructive suggestions. We answer each question below. All success rates are computed over 15 checkpoints: the last 5 epochs of each of 3 training seeds.

W1/Q1. Comparison with relative translation actions.

We agree that a fixed Gaussian on relative translations Δp=pa−pee\Delta p=p_a-p_{ee} is compatible and mathematically equivalent to our anchored source N(pee,σ2I)\mathcal{N}(p_{ee},\sigma^2I) under the change of variables pa=pee+Δpp_a=p_{ee}+\Delta p. However, the two formulations differ in how actions are represented within the network. In GARLIC, absolute actions are embedded as points, which transform under both translations and rotations. In the relative-action formulation, displacements are embedded as directions, which are invariant to translation and equivariant to rotation.

To evaluate the practical implications of this distinction, we implemented the requested relative-action baseline using the same GATr backbone. We evaluate on Unplug Charger with 100 demonstrations. The table reports in-distribution (ID) and out-of-distribution (OOD) success rates.

MethodAction representationSourceID (%)OOD (%)
Relative-actionRelative displacement (direction)N(0,σ2I)\mathcal{N}(0,\sigma^2I)65.5±3.564.7±5.7
GARLICAbsolute position (point)N(pee,σ2I)\mathcal{N}(p_{ee},\sigma^2I)84.7±5.183.2±4.7

Both formulations exhibit small ID–OOD performance gaps, consistent with their source compatibility. However, GARLIC achieves substantially higher success rates, and also achieves lower validation translation error (13.6 mm vs 32.2 mm).

We attribute this to the action embedding. With absolute actions, the action is a point in the scene, so its spatial relation to the scene is directly available to the network. With relative actions, the network sees only a direction and must learn to compose the target position internally. This matches the analysis in de Haan et al. (AISTATS 2024).

These results show that the anchored source preserves compatibility while letting actions remain points in world coordinates, which substantially improves performance with the GATr backbone. Compatibility therefore no longer requires switching to relative actions, and the action representation can be selected on empirical grounds.

We added this comparison as Section 6.5, with implementation details in Appendix D.3.

Q2. Compatibility versus proximity to target actions.

We agree that anchoring the source at the end-effector also moves the source closer to the demonstrated actions. We compare GARLIC with 3 alternative source distributions:

  • Shifted-anchor: N(pee+0.5Reeez,σ2I)\mathcal{N}(p_{ee}+0.5R_{ee}e_z,\sigma^2I), a compatible source whose center is shifted 0.5 m from the end-effector along its local z-axis.
  • Mean-centered: N(pˉ,σ2I)\mathcal{N}(\bar p,\sigma^2I), the observation-independent Gaussian suggested by the reviewer, centered at the mean training action pˉ\bar p.
  • Fixed Gaussian: N(0,σ2I)\mathcal{N}(0,\sigma^2I), the observation-independent source used in the original ablation study (Table 4).

We measure proximity as the mean distance between each source center and the corresponding demonstrated action position. Δ=OOD−ID\Delta=\mathrm{OOD}-\mathrm{ID}.

SourceCompatibilityDistance (m)IDOODΔ\Delta
GARLIC✓0.23084.7±5.183.2±4.7−1.5
Shifted-anchor✓0.42982.4±1.482.7±2.8+0.3
Mean-centered✗0.20574.2±2.763.4±6.9−10.8
Fixed Gaussian✗0.47550.7±5.531.3±5.0−19.4

Source-to-target proximity alone does not explain the performance differences. At comparable distances, the compatible sources outperform the incompatible ones. GARLIC outperforms Mean-centered by 10.5 points in ID and 19.8 in OOD despite a larger distance, and Shifted-anchor outperforms Fixed Gaussian by 31.7 and 51.4 points despite their comparable distances. They also show smaller ID–OOD gaps, consistent with the benefits of equivariance.

The effect of source placement depends on compatibility. Among the compatible sources, increasing the distance from 0.230 m (GARLIC) to 0.429 m (Shifted-anchor) reduces ID success by only 2.3 points. In contrast, moving the incompatible Fixed Gaussian from the origin to the mean training action reduces its distance from 0.475 m to 0.205 m and improves ID success from 50.7% to 74.2%. This is consistent with the reviewer's observation that a more favorable source distribution can facilitate learning, as also suggested by prior work on informative sources (Lee et al., ICLR 2022; Chen et al., RSS 2024).

Together, these results show that compatibility is the key factor, and that source placement matters mostly when compatibility is absent. Once the source is compatible, its placement becomes a free design choice with little effect on performance. GARLIC anchors it at the end-effector because this point is directly observed.

We added this comparison as Section 6.6, with details in Appendix D.4.

Q3. OOD evaluation protocol and real-world deployment.

In our simulated OOD evaluation, we apply planar rigid transformations to the entire scene, including the task objects, robot base, arm, and initial end-effector pose. This ensures that the end-effector position, which anchors GARLIC's translational source, transforms consistently with the scene. The protocol follows EquiBot (Yang et al., CoRL 2024), whose simulated evaluation transforms the robots together with the objects, and corresponds to a manipulator on a mobile base. The observations are obtained from four fixed external cameras and one wrist-mounted camera that moves with the robot.

Because the external cameras remain fixed, the resulting point clouds are not exact rigid transformations of the original observations. Changes in viewpoint introduce differences in visibility and occlusion. Thus, the OOD experiments evaluate policy performance under joint scene transformations and the accompanying changes in sensory observations, rather than exact equivariance under transformed point clouds. GARLIC maintains OOD success rates within 2 percentage points of its ID performance across all four tasks.

We agree that the theoretical equivariance guarantee does not extend to object-only transformations with a fixed robot base, since these are not a rigid transformation of the complete observation. This setting is evaluated empirically: the ID experiments randomize object configurations with a fixed base (Table 2), and the real-robot experiments use a fixed base (Table 5). They show that GARLIC learns to manipulate objects at different positions without moving the robot base.

Evaluating on a real mobile manipulator, and under more systematic object-only shifts, is left for future work. We added this to the conclusion of the manuscript.

Weakness: contributions of the source design and GATr backbone.

Both the source design and GATr contribute to the performance gains, and each is isolated by a controlled comparison.

Effect of the source distribution (A1). Holding the GATr backbone fixed, replacing the anchored source with Mean-centered reduces ID success from 84.7% to 74.2% and OOD success from 83.2% to 63.4% (Q2). Replacing it with the Fixed Gaussian further reduces performance to 50.7% ID and 31.3% OOD (Table 4). Both alternatives violate source compatibility under translations (A1). As discussed in Q2, these performance differences cannot be explained by source-to-target proximity alone.

Effect of the backbone (A2). Holding the anchored source fixed, replacing GATr with the non-equivariant PFM backbone reduces ID success from 84.7% to 61.7% and OOD success from 83.2% to 8.7% (Table 4, "Remove GATr"). This demonstrates the substantial practical contribution of the GATr backbone, including its equivariant architecture, to GARLIC's OOD performance.

These ablations mirror the theory in Section 3. Source compatibility (A1) and an equivariant generation mechanism (A2) together guarantee an equivariant policy, and neither condition alone does. Accordingly, removing either one lowers performance, particularly under the evaluated OOD transformations.

Thank you for the detailed response. The rebuttal has addressed several of my concerns, and I will maintain my positive rating.

Reviewer_7RhWOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper shows that an equivariant generative robot policy requires a compatible source distribution, and that invertible equivariant flows preserve source equivariance errors. It proposes GARLIC, combining end-effector-anchored noise, Haar-distributed rotations, and an equivariant Transformer. A controlled planar study validates the theory, while simulated and real-robot tasks demonstrate improved manipulation performance, with simulated evaluations showing robustness to scene transformations.

Critical strengths

A clear analysis of source-induced equivariance error. The paper explains why invertible equivariant flows cannot correct incompatible sources and provides practical guidance for source design. A controlled planar experiment supports the theoretical predictions. Due to time constraints, I did not verify mathematical proofs in detail.

Critical weaknesses

The statistical reporting seems inconsistent. Section 6.2 states that checkpoint results are averaged within each seed before reporting variability across seeds, whereas Appendix D.2 describes reporting variability across checkpoint success rates. These procedures measure different sources of variation, and checkpoints from the same training run are correlated.

Questions for authors

Please clarify the weakness

1 comment

W1. Statistical reporting across checkpoints and seeds

Thank you for identifying this inconsistency. Section 6.2 incorrectly describes the originally reported standard deviations as being computed across training seeds, whereas Appendix D.2 correctly describes the original calculation across checkpoints.

Following EquiBot (Yang et al., 2024), we train 3 seeds and evaluate the last 5 checkpoints of each run. Each checkpoint is evaluated on the same 100 randomized episodes, and we report the mean and standard deviation of the success rate over the 15 checkpoints.

We agree that checkpoints from the same run can be correlated. To isolate the seed-to-seed variation, we therefore recomputed the GARLIC results in Table 2 by first averaging the success rates across the five checkpoints within each seed and then computing the standard deviation across the three resulting seed-level means.

The table below compares the original and recomputed results, reported as mean success rates ± standard deviation (%).

# DemosStd overTake lid off saucepanWater plantsUnplug chargerTake frame off hanger
1015 checkpoints (paper)40.7 ± 5.542.1 ± 6.233.7 ± 3.138.7 ± 4.2
3 seeds40.7 ± 6.042.1 ± 6.733.7 ± 3.438.7 ± 4.8
5015 checkpoints (paper)92.0 ± 2.481.3 ± 1.556.5 ± 3.463.7 ± 2.9
3 seeds92.0 ± 2.881.3 ± 1.756.5 ± 3.763.7 ± 3.2
10015 checkpoints (paper)95.7 ± 1.790.7 ± 2.185.7 ± 5.271.3 ± 4.0
3 seeds95.7 ± 1.690.7 ± 2.285.7 ± 4.971.3 ± 3.6

The recomputed standard deviations differ from the original values by at most 0.6 percentage points, while the mean success rates remain unchanged. Thus, the corrected aggregation does not alter GARLIC's reported mean performance or its ranking across the evaluated tasks and demonstration budgets.

We have revised Section 6.2 accordingly in the manuscript. Thank you again for catching this inconsistency and for your positive assessment of our paper.

Comments on the paper

1 comment

We thank all three reviewers for their thoughtful feedback and constructive suggestions. We believe the additional experiments and manuscript revisions substantially strengthen the paper.

We are encouraged that the reviewers found our paper to offer a clear and useful framework for analyzing how the source affects equivariance (7RhW, eGSi, HnKM); be well motivated and address an important question that has received little attention in equivariant robot policies (eGSi, HnKM); explain why invertible equivariant flows cannot correct incompatible sources (7RhW); support its theoretical predictions with a controlled planar experiment (7RhW, eGSi); and provide practical guidance for source design (7RhW).

Below we summarize the main concerns raised by the reviewers and how we addressed them.

Source compatibility versus proximity. To separate the effect of compatibility from that of placing the source closer to the target actions (eGSi, HnKM), we added sources that vary in compatibility and in distance (Section 6.6, Table 7, Appendix D.4). The table below summarizes our experiments, where columns give the approximate mean distance from the source center to the target actions and each cell reports the source and its ID / OOD success (%).

Distance ≈ 0.2 mDistance ≈ 0.45 m
CompatibleGARLIC: 84.7 / 83.2Shifted-anchor: 82.4 / 82.7
IncompatibleMean-centered: 74.2 / 63.4Fixed Gaussian: 50.7 / 31.3

These results show that compatibility is the key factor, especially under OOD transformations: at comparable distances, the compatible sources substantially outperform the incompatible ones and show smaller ID–OOD gaps. The effect of proximity depends on compatibility, and it matters mainly when compatibility is absent.

Relative translation actions. To compare absolute actions with the anchored source and relative actions with a fixed Gaussian in practice (eGSi), we added the requested relative-action baseline using the same GATr backbone (Section 6.5, Table 6, Appendix D.3). Both formulations are compatible and show small ID–OOD gaps, but GARLIC achieves higher success (84.7 / 83.2 vs. 65.5 / 64.7) and lower validation translation error (13.6 mm vs. 32.2 mm). We attribute this to the action embedding: absolute actions are embedded as points, which are equivariant to translation, whereas relative actions are embedded as directions, which are invariant to translation, so their spatial relation to the scene is not directly available to the network.

End-effector-frame canonicalization. To compare with canonicalization using a matched architecture (HnKM), we evaluated frames anchored at the end-effector with world, end-effector, and random axes, on both GATr and DP3 (Section 6.7, Table 8, Appendix D.5). With the equivariant GATr backbone, the three orientations perform comparably, differing by at most 1.5 points in both ID and OOD. Explaining this led us to formalize a property that is usually used only implicitly: with an equivariant backbone, canonicalization requires an anchor point but no frame orientation (Remark 2, Appendix A.6). With the non-equivariant DP3 backbone, the frame orientation matters.

Clarifications. We clarified that the OOD protocol applies planar rigid transformations to the entire scene including the robot (Appendix D.2), while the external cameras remain fixed, so the observations change in visibility and occlusion (eGSi). We revised the abstract and conclusion to state the scope of our evaluations explicitly (eGSi, HnKM). We corrected the description of the statistical reporting in Section 6.2 (7RhW). We also recomputed the standard deviations across seeds to isolate the seed-to-seed variation from correlations within a training run.

The revised manuscript incorporates the reviewers' feedback, with all changes highlighted in blue, and each concern is addressed in detail in the individual replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.