acceptodds
Withdrawn from ICLR 2027

Geometry Beyond Symmetry: On-Manifold Equivariant Graph Networks for Robot Dynamics

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Predictive controllers rely on dynamics models to evaluate candidate action sequences, making accurate multi-step pose prediction critical to reliable planning and control. Robot dynamics present two intertwined structures: their predictions should remain consistent across equivalent reference frames, while their pose states evolve on a non-Euclidean configuration space. Existing -equivariant graph neural networks encode symmetry for Euclidean point coordinates, but do not respect the manifold geometry of rotational states. We introduce an On-Manifold Equivariant Graph Neural Network (OM-EGNN) that represents robot poses on the product manifold of translation and rotation, computes relative poses in receiver-local tangent coordinates, and retracts learned updates onto the manifold. Moreover, the radial update of standard EGNNs is restricted to scalar-weighted combinations of pairwise displacements. OM-EGNN augments this update with a learned tangent-space residual that escapes this directional span while preserving task-relevant equivariance. On planar Ackermann-steered vehicle dynamics and spatial quadrotor dynamics, we compare OM-EGNN against standard canonicalized predictors and Euclidean equivariance-based graph baselines. OM-EGNN achieves the lowest position RMSE in both settings. Rotation and velocity are not uniformly best: planar velocity is lowest for a Euclidean lift, and 3D rotation is lowest for a canonicalized MLP. Unlike the Euclidean graph baselines, OM-EGNN keeps the frame-attack gap near zero under the task symmetry. The tangent residual substantially improves position prediction. In matched MPPI, OM-EGNN has the best mean progress, success, and tracking RMSE, but those margins lie within run variation of the strongest canonicalized predictors, and open-loop rankings do not determine the control ranking.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Equivariant Learning, Manifold, Robot dynamics learning

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_LP5EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes an approach to using a graph neural network to reason about dynamics embedded in SE(3). The paper explores an equivariant model that accomplishes the above. This is in contrast to most approaches in the literature that reason about SO(3) as embedded in a Euclidean vector space where, e.g. quaternions are treated as a vector quantity. Essentially, the proposed method reasons in exponential coordinates and projects the resulting orientation back onto the SO(3) manifold. The work is evaluated in the context of a forward dynamics model that predicts future states. Table 1 and Fig 2 evaluate performance of the model in terms of predictive error for a planar ackerman dynamics vehicle. Table 4 evaluates planning performance for MPPI using this learned model for the ackerman vehicle.

Critical strengths

The main strength of the paper was its mathematical soundness. While the tangent map introduces an approximation, I found the method generally well founded. In addition, I thought the GNN framework was appropriate.

Critical weaknesses

My main concerns relate to the experiments. The problem was that they focused primarily on the dynamics of the planar vehicle. It's just not a problem that really requires this kind of machinery. Plus, it's a problem in SE(2), not SE(3). The strength of the math is really SE(3), but that's not really needed for these experiments. I cannot believe that really good approaches to planar vehicle dymamics prediction do not exist that are significantly simpler than this.

Questions for authors

No questions. My primary concern relates to the experiments.

Reviewer_di4ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes OM-EGNN for direct multi-step prediction of robot dynamics. The model represents poses on the product manifold of translation and rotation and constructs a causal graph containing observed history and future query nodes. Relative poses are expressed in receiver-local tangent coordinates, and learned increments are retracted onto the pose manifold to preserve valid rotations and task-relevant equivariance. A learned tangent-space residual supplements the scalar-weighted displacement update, addressing an expressivity limitation that prevents displacement-only models from predicting motion when all initial poses coincide.

The method is evaluated on simulated planar vehicle dynamics and offline quadrotor flight data from NeuroBEM, with additional planar friction holdouts, reference-frame diagnostics, and matched MPPI experiments in simulation. OM-EGNN achieves the lowest reported position errors in the main prediction comparisons, while rotation and velocity results are mixed. Its closed-loop performance is competitive with strong canonicalized predictors, with relatively small differences between them.

Critical strengths

The architectural motivation is clear and supported by a concrete expressivity argument. The paper distinguishes reference-frame consistency, valid rotational updates, and the ability to generate control-dependent motion. In particular, the analysis of displacement-only updates at rest provides an understandable explanation for the tangent residual, rather than introducing it solely as an empirical modification. The receiver-local construction also makes the equivariance argument transparent.The empirical evaluation includes strong canonicalized predictors and examines both prediction and downstream control. The position-prediction improvements in Tables 1 and 2, together with the trajectory-disjoint planar friction holdouts, provide useful evidence for the proposed model. I also appreciate that the paper explicitly acknowledges mixed rotation and velocity results and the modest control margins over canonicalized baselines.

Critical weaknesses

The paper does not yet provide sufficiently strong evidence of substantial benefits in practical control settings. Closed-loop evaluation is limited to a planar simulation environment, while the NeuroBEM experiments use real flight logs for offline prediction and do not involve deploying the model on physical hardware for closed-loop control.

Given the paper’s predictive-control motivation, it remains unclear whether the proposed architecture offers meaningful practical advantages under real-world dynamics, sensing errors, and actuation delays. Although the authors acknowledge this limitation, the modest control gains in simulation, together with the lack of hardware validation, leave me unconvinced that the practical contribution has been sufficiently established.

I also suggest that the authors review the manuscript’s formatting.

Questions for authors

Would it be feasible to provide a focused closed-loop hardware comparison between OM-EGNN and the strongest canonicalized baseline, using the same controller and comparable inference budgets? A small-scale planar vehicle experiment would already be informative. The main question is whether the predictive improvements translate into a meaningful control benefit on a physical system, given the relatively small margins in the simulated MPPI results. Such evidence would materially affect my assessment of the paper’s practical contribution.

Reviewer_ssJnOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes OM-EGNN, an equivariant GNN for learning robotic predictive control. Specifically, OM-EGNN predicts multi-step future robot state from state-control history and future controls. The authors argue that existing equivariant GNNs treat state geometry as Euclidean points, which is not manifold-aware and restricts the expressivity of the internal predictions. OM-EGNN solves this issue by updating internal nodes with tangent vectors that are canonicalized in a receiver-local manner. Experiments on a Ackerman steering vehicle and quadrotor data show that the proposed method outperforms existing equivariant GNN baselines in terms of accuracy.

Critical strengths

The authors openly report cases where baselines outperform their method and acknowledge the limitations of their claims. Results are reported over five seeds with standard deviations, and code and checkpoints are provided in the supplementary material.

Critical weaknesses

[Concern 1] The authors argue that existing methods are not manifold-aware, which is a problem.

Existing E(n)-equivariant graph neural networks encode symmetry for Euclidean point coordinates, but do not respect the manifold geometry of rotational states.

However, I do not agree with this claim. While a single Euclidean point can't encode a rigid body, n+1n+1 non colinear points can faithfully encode E(n)E(n). Moreover, equivariant GNNs such as TFN and SE(3)-transformers can encode a rigid body with a single point by leveraging higher-order steerable features. A simpler solution is to attach a fixed learnable point cloud at each reference frame, and use existing EGNNs. While the authors show that the span of the EGNN's internal point update can be degenerate in certain cases, but this is a limitation of EGNN, not a general limitation shared across different equivariant architectures.

[Concern 2] The scope of the proposed method is narrow, and the experimental evaluation is limited. The experiments are conducted in very small scale only (about 200k param networks) and thus, OM-EGNN's scalability is not verified. As a result, the broader relevance of the contribution to the ICLR audience is unclear.

Questions for authors

Is it possible to disentangle the effects of on-manifold updates and architectural equivariance? Without controlled experiments, it is difficult to attribute the reported gains to either factor.

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

Related papers

Discussion (0)

Sign in to comment.