acceptodds
Under review as a conference paper at ICLR 2027

Equivariant Lagrangian Neural Networks with Noether Guarantees

Abstract

Learning physical dynamics from data supports simulation and control. Lagrangian Neural Networks (LNNs) learn a scalar Lagrangian in generalized coordinates and derive dynamics through the Euler-Lagrange equations, but standard LNNs do not guarantee spatial symmetries or their conservation laws. We introduce Equivariant Lagrangian Neural Networks (EqLNNs), which lift nonlinear generalized coordinates and velocities to features with known linear transformation laws and use an equivariant network to produce an invariant scalar Lagrangian. We show that enforcing invariance in the lifted representation makes the resulting Lagrangian invariant to the corresponding transformations of the original generalized coordinates. By Noether’s theorem, this guarantees the corresponding conserved quantities when the learned Lagrangian is regular. The construction handles translations, articulated and rigid-body systems, and fixed external fields through their remaining symmetries. Across five systems, EqLNN improves over a lifted LNN on four systems, including to times lower median trajectory MAE on the free-body systems. Momentum components tied to enforced symmetries drift by at most , while matched lifted LNNs show order-one drift in the corresponding angular-momentum components. EqLNN can improve prediction while making the appropriate conservation laws part of the model rather than something inferred from data.

Primary area:
applications to robotics, autonomy, planning
Keywords:
equivariant neural networks, Lagrangian neural networks, Noether's theorem, geometric deep learning, conservation laws, physics-informed machine learning, generalized coordinates, rigid-body dynamics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SYWsOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors aim to learn dynamical systems with certain invariant properties. They introduce the Equivariant Lagrangian Neural Network (Eq-LNN). The Eq-LNN lifts the space to one where features transform linearly. They use an equivariant neural network architecture which outputs a scalar Lagrangian that is invariant by design, ensuring that the desired invariances and, by Noether's theorem, conserved quantities of the original system are preserved. The network can also handle adjustments for gravity and translations. The method is demonstrated on multiple systems with different types of invariances, to strong results.

Critical strengths

S1. The authors provide a novel approach to a well-defined problem, learning dynamical systems with invariant properties.

S2. The approach has strong mathematical motivation.

S3. The proposed method performs extremely well on most systems.

Critical weaknesses

W1. The negative results of the coupled pendulum and spaceraftA are not discussed.

W2. The most recent baseline is from 2019 (Lutter et al.) and, in the appendix, 2020 (Finzi et al.). More recent work has occurred in this domain, as discussed in the Related Work section of this paper.

W3. The Related Work section could do a better job of framing the paper. The section as it stands reads like a bullet point list, without providing an overall framework for deciding when and why to use your method. It should be clear from this section and the introduction why the more recent work was not included in the comparison.

Questions for authors

Q1. Why don’t you include more recent work in your baselines? The lack of more recent baselines suggest that no progress in the field has been made since 2019 (2020 in the appendix). As stated above, this should be answered by a combination of the introduction and related work sections. Q1b: Why was CLNN compared against only in the appendix?

Q2. What physical properties of the chained pendulum make the EQ-LNN perform as well as the Lifted LNN? Your results are very impressive as they are. It is not required to win in all cases. But, answering this question would greatly help understand the properties of your architecture and make for a strong paper.

Same goes for the lift comparison with SpacecraftA. “Equivariance cannot recover geometry discarded by the lift” is given from the result. a physics informed explanation would be very valuable.

Q3. What are the computational tradeoffs? It seems like some of the architecture is computationally costly. I would like to see a comparison on runtime with prior work. Its possible that the cost is negligible in absolute time. I would also be interested in understanding the effect of the rank r on timing and performance.

Q4 A lot of things about the system have to be known in advance. Each system appears to require custom design choices including: the position and velocity lifts, the existence of global translations, and the presence of external forces (gravity). In what context would I know all of these properties of the system but still need machine learning to learn the system?

2 comments

First, we would like to thank the reviewer for their time, effort and thoughtful comments. We appreciate that you are probably very busy and took time out of our day to perform this review. Below we provide some additional experiments, as well as clarifications and comments

W1 — CoupledPendulum, ChainPendulum, and SpacecraftA.

For CoupledPendulum, including previously omitted anchor-direction context improves EQ-LNN from 4.68e-2 to 4.22e-2, although Lifted LNN remains slightly better (3.98e-2). Fixed anchors and gravity leave no continuous spatial symmetry to enforce. Equivariant restrictions may therefore complicate learning without providing a corresponding conservation benefit, but the ~6% gap could also reflect seed variation.

SpacecraftA For SpacecraftA, the original minimal lift supplied the main body's orientation and the two appendage hinge angles, but did not expose how the appendages move relative to the main body. In particular, it omitted the hinge locations, the offsets of points along the appendages, and the resulting point velocities.

For the rebuttal, we extended our minimal-lift recipe with body attached points: virtual reference locations attached to each rigid component. three shape vectors in the bus frame, plus a hinge offset and one marked point per appendage. A generic kinematic operator then composes these with the bus attitude and hinge rotations to get world-frame landmark positions and their velocities. This avoids the system-specific jacobian derivation, and only requires coarse geometry.

This lift has 45 features against 69 for the analytic lift, and no derived Jacobian. With this update, minimial lift Eq-LNN gets back to being close to the analytic lift Eq-LNN: 4.14e-4 median trajectory MAE, versus 2.11e-2 for the matched Lifted LNN on identical features.

Q2

On ChainPendulum, EQ-LNN does not perform comparable to Lifted LNN. It is 18.2x better than Lifted LNN (9.02e-4 vs. 1.64e-2).

W2–W3 / Q1 — Recent baselines and related-work framing.

We originally prioritized Lifted LNN to isolate the architecture with identical inputs and training, but inadquately explained the absence of newer comparisons. We now add invariant-scalar LNNs based on Müller (2023), our separate lift-scalarization extension, and Liu et al.'s GHNN (2025). Note that:

Müller Scalarization: Constructs invariant scalars directly from planar particle positions and velocities, following Müller (2023). Muller's contribution focused on particle type dynamics and didn't apply to rigid body experiments.

Lift + scalar (ours): Our new extension that first applies our symmetry-compatible lift and then scalarizes those lifted features for a dense LNN. This was not proposed or evaluated by Müller.

SystemEQ-LNNLifted LNNMüllerLift + scalar (ours)GHNN '25
Two-body1.77e-24.26e-22.95e-3--2.49e-1
TwoBodySpring7.90e-54.14e-38.61e-47.06e-42.27e-1
ChainPendulum9.02e-41.64e-2----1.89
CoupledPendulum4.68e-23.91e-2----2.04
Gyroscope4.95e-31.92e-1--1.10e-32.15
SpacecraftA*4.14e-42.11e-2----2.66e-1
SpringGraph(8), chain4.52e-42.51e-23.45e-21.88e-25.34e-1

Scalarization wins on Two-body and Gyroscope; EQ-LNN has 17–44x lower error than the best scalar baseline across SpringGraph chains (N=2,4,8,16). SpacecraftA* uses the updated minimal lift only for the matched EQ/Lifted pair. Dashes mean unrun.

GHNN discussion GHNN (Liu et al., 2025) is a relevant recent baseline because it does not require canonical position–momentum coordinates (unlike HNN). Instead, it learns a Hamiltonian and a state-dependent interconnection matrix, allowing dynamics to be modeled directly from noncanonical state representations and accommodating constrained systems.

We could not locate an official public implementation for GHNN, so we implemented it using their descriptions and evaluated using our common benchmark protocol. We attempted to identify why it lost performance on the pendulum systems, with several diagnostic investigaiton, but htose were inconclusive. One possible explanation is that learning the interconnection structure introduces additional optimization difficulty. We regard these results as specific to our benchmarks and implementation of GHNN, not as evidence against GHNN's broader applicability.

We will organize Related Work by what is supplied: invariant scalars, equivariant features, graph structure, explicit constraints, or Hamiltonian structure. Recent graph methods address different architectural and transfer questions.

Q1b CLNN discussion

We placed CLNN in the appendix because it is not a one-to-one comparison with Eq-LNN, and unfortunately we didn't have space to fully explain in the main paper. CLNN is making different assumptions, and is closer to structured system identification.

CLNN is closer to structured system identification: it assumes known bodies, joints, and constraint equations (including jacobians). This additional requirement for a user-supplied mechanical formulation leaves CLNN responsible for learning only a static, configuration-independent Cartesian mass matrix on the kinetic-energy side. This greatly simplifies the learning problem: instead of learning the complex configuration-dependent kinetic energy structure, CLNN only learns/identifies constant mass and inertia parameters. (For potential energy CLNN separately learns it through a neural network, which is relatively simple for several of these benchmarks). Much of the complexity is therefore handled by the supplied mechanics and constraint projection rather than learned by the neural network.

In contrast, EQ-LNN learns a general invariant Lagrangian, including its kinetic-energy dependence, directly in generalized coordinates without prescribing a fixed quadratic kinetic-energy form or requiring explicit constraint projection. Thus, CLNN comparison is informative, but comparing a different type of algorithm. We will update our text with this clarification, and also clarify this distinction in related work.

Q3 — Runtime and rank.

We ran a controlled SpringGraph(8) comparison. EQ-LNN is smaller (77k vs. 140k parameters) but slower: 5.98 vs. 0.79 s per training step, and 1.44 vs. 0.21 s per evaluation trajectory. That ~7x runtime cost is a limitation.

We also compared the rank of the bilinear factorization. On ChainPendulum(2), ranks 4/8/16 give MAE 8.72e-4 / 7.83e-4 / 5.75e-4 at about the same cost (1.77–1.80 s/step).

Q4 — What must be known, and when is learning still necessary?

Our setting assumes the system's symmetries and meaning of position/velocity inputs are known, but the governing dynamics are not. These are very different levels of physical knowledge: knowing how a system transforms under rotations or translations does not determine its masses, moments of inertia, interaction strengths, or potential energy.

The lifts are built from reusable primitives, not a custom derivation for every system. Our minimal-lift recipe uses simple components with known transformation properties: joint angles become unit-direction vectors, rigid-body orientations become frame axes, and more complicated articulated systems use a sparse set of marked points. Differentiating these maps gives the velocity features. The input representation does still have to be specified, but no hand-derived Lagrangian, mass matrix, or equations of motion are required.

The remaining choices, translation and gravity, are also physically interpretable. Whether a system is free to translate or attached to fixed anchors is generally known from its setup. Gravity's direction can also be known without knowing the system's gravitational potential. Our gravity branch uses this direction and the corresponding position features, but learns the potential instead of prescribing its functional form or coefficients.

As an example, take a spacecraft with known joint structure and approximate geometry, but uncertain mass distribution, payload, inertial properties, or conservative interactions. Those quantities can affect the dynamics substantially and may be difficult to measure or model analytically. EQ-LNN can learn their effect directly from trajectory data while guaranteeing consistency with the known symmetries.

So we target learning unknown dynamics from known structural information, not discovering symmetries or geometry from scratch. That is a deliberate modeling assumption, and also a common setting where physics-informed learning is useful.

Thanks again for your time and feedback. Let us know if there are additional clarifications or experiments we can add.

Reviewer_BU4WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes EQ-LNN, which combines equivariant coordinate lifts with an invariant network to learn Lagrangian dynamics. It derives Noether conservation guarantees under explicit regularity assumptions and evaluates predictive accuracy and conservation behavior on particle, articulated-body, and rigid-body systems.

Critical strengths
  1. The framework connects architectural symmetry constraints with conservation properties of the continuous learned dynamics.

  2. The matched lifted-LNN comparison is useful, with substantial improvements on four of the five primary systems (Table 1).

  3. The conservation experiments distinguish preserved and broken symmetry directions, providing a useful check of the intended behavior (Figure 3).

Critical weaknesses
  1. The experiments lack systematic comparisons across different body counts and interaction topologies, leaving applicability to larger and more complex many-body systems unclear.

  2. For CoupledPendulum, the relative-coordinate lift and invariant network appear to impose rotational invariance that may be inconsistent with the fixed anchors and spring couplings. This potential mismatch needs clarification.

Questions for authors

See above.

1 comment

We would like to thank the reviewer for their thoughtful comments and feedback. We added topology experiments and tested the coupled pendulum including the anchor offsets.

Body count and topology

To test different body counts and topologies we used our two body spring system and used it to construct various graph based topologies: including a sequential chain, a ring, and star shaped topologies. We also tested scaling up the number of bodies.

NTopologyEQ-LNNLifted LNNGain
2Chain5.76e-53.60e-362.5x
4Chain2.07e-41.53e-273.8x
8Chain4.52e-42.51e-255.6x
8Ring4.29e-42.35e-254.8x
8Star5.19e-43.35e-264.4x
16Chain1.63e-33.22e-219.8x

EQ-LNN's learned momentum drift stays at numerical tolerance in every system test. Each system was trained separately, so these are showing performance across sizes and topologies, not zero-shot size or topology transfer, and we will state that explicitly.

CoupledPendulum anchors

The reviewer is also correct about CoupledPendulum. The original invariant branch used bob directions but omitted the fixed anchor offsets, which are needed to represent the directional spring interactions.

We have now added the centered anchor offsets as configuration-only covariant features. This does not imply that the physical system has rotational symmetry: the anchor positions and gravity direction are fixed, so rotating the pendulums alone changes the system's potential energy. We therefore make no spatial Noether conservation claim for CoupledPendulum.

CoupledPendulumEQ-LNNLifted LNN
Original lift4.68e-23.91e-2
+ anchor context4.22e-23.98e-2

Adding the anchor information reduces EQ-LNN's median error by about 10%. The correction helps, but Lifted LNN is still slightly more accurate on CoupledPendulum. Both now receive the same information. The remaining small difference in performance may be due to different optimization landscapes due to the equivariant basis construction or the factorization of teh bilinear layers, and could potentially reflect seed variability.

Spacecraft

Our initial submission included several spacecraft systems in the appendix. SpacecraftA is a rigid bus with two hinged appendages (translation, MRP attitude, two hinge angles; 8 generalized coordinates). In the initial submission we noted that our full analytic lift succeeded on SpacecraftA while our minimal lift struggled. During the rebuttal we investigated this limitation and improved our minimal lift recipe by incorporating sparse marked-point kinematics.

Instead of the full analytic Cartesian geometry, the recipe asks for a handful of body-attached points: three shape vectors in the bus frame, plus a hinge offset and one marked point per appendage. A generic kinematic operator then composes these with the bus attitude and hinge rotations to get world-frame landmark positions and their velocities. This avoids the system-specific jacobian derivation, and only requires coarse geometry.

This lift has 45 features against 69 for the analytic lift, and no derived Jacobian. Even with fewer features, EQ-LNN gets back close to the analytic result: 4.14e-4 median trajectory MAE, versus 2.11e-2 for the matched Lifted LNN on identical features.

We thank the reviewer again for their feedback. We feel it's led to strengthening the work. We hope we were able to address these concerns.

Reviewer_7ChGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an LNN that imposes spatial symmetries using lifted features and an equivariant network. The construction retains generalized coordinates and gives Noether conservation guarantees for the regular learned dynamics. It is tested on synthetic mechanical systems.

Critical strengths

I think retaining generalized coordinates while imposing known symmetries is useful. I also appreciate the comparisons using the same lifted inputs and multiple training seeds.

Critical weaknesses

My main concern is that I am not convinced there is enough of an advance over Müller (2023), which already constructs invariant neural Lagrangians with Noether conservation guarantees. I think this paper explores a broader approach in some ways, but it does not do enough comparisons against these previous approaches. Without that comparison, I think we cannot tell what the proposed architecture adds over the simpler approach. Also, I found the examples were unconvincing in that they were too small and noiseless to really act as a proxy for any practical benefits on real problems. So I think the contribution needs stronger experimental support.

Questions for authors

How does EQ-LNN compare with a full LNN using invariant scalar features under the same training protocol?

4 comments

We thank the reviewer for their time and thoughtful comments. We agree that the original submission did not sufficiently describe Müller's method or its relation to our work, and we have added direct scalarization comparisons and larger-system experiments to address this concern.

What differs between Müller et al. and our approach

In Müller et al., the invariant scalar features are designed by hand for each system and fed to a standard LNN — feature engineering, as their repository describes it. The features they publish are already close to the exact Lagrangian: Kepler reduces to a function of squared speed and squared distance, and their two-particle example is a quadratic in squared separation and squared speeds (Eqs. 13, 22–23). So the hard geometric work is done before learning starts, and the network only has to fit a low-dimensional scalar relationship.

We differ in two ways:

  1. The lift. Most of our benchmarks are in generalized coordinates — joint angles, rigid-body attitudes, hinge phases — not Cartesian particle states, so there is no obvious set of invariant scalars to feed an LNN. Our minimal lift builds vector features from these coordinates and their tangent velocities; without it, invariant Lagrangian learning does not apply to these systems at all.

  2. The equivariant neural network: Eq-LNN's bilinear layers progressively construct interactions using intermediate equivariant features. In contrast, scalarization computes a fixed set of scalar inputs. Because bilinear layers pass representation-valued features forward, they can also construct higher-order invariants when the relevant compositions are not known in advance.

The other problem with scalarization is size. Twelve SO(3) vectors already give 78 pairwise dot products and 220 oriented triple products, and the count grows fast from there. The bilinear layers avoid constructing this set at the input altogether.

Our framework also handles uniform-direction external gravity through the position-only branch, preserving the residual symmetries. Müller's work included central gravitational wells acting on a particle, but not fixed-direction external forces (such as gravity on Earth). Their approach could likely be extended to such settings, but they do not demonstrate this.

Comparison with Scalarization

To evaluate the differences, we tested two scalarization baselines:

Müller-style scalarization: using Müller's method for exhaustive pairwise dot products of the standard vectors (relative positions, absolute velocities), fed to the same dense backbone. Lift + scalarization (ours): the same scalarization applied to channels produced by our lift. This was required for rigid-body systems such as Gyroscope.

Table 1: Invariant-scalar LNN comparison. Median trajectory MAE; lower is better.

SystemEQ-LNNLifted LNNMüller-style scalarizationLift + scalarization (ours)
Two-body1.77e-24.26e-22.95e-3--
TwoBodySpring7.90e-54.14e-38.61e-47.06e-4
Gyroscope4.95e-31.92e-1--1.10e-3
Rotor1.81e-36.73e-2--Failed

Scalarization is effective when a sufficient invariant feature set is available. Where the set grows large, EQ-LNN wins instead, as it does on TwoBodySpring (8.9x below the best scalarized arm). On Rotor the SO(3)-scalarized head failed to train (three non-finite seeds, plus dot-only and lower-lr diagnostics).

Table 2: SpringGraph scaling and topology. Median trajectory MAE; lower is better.

NTopologyEQ-LNNLifted LNNMüller-styleLift + scalarization (ours)
2Chain5.76e-53.60e-31.16e-39.90e-4
4Chain2.07e-41.53e-21.75e-29.12e-3
8Chain4.52e-42.51e-23.45e-21.88e-2
16Chain1.63e-33.22e-25.38e-22.84e-2

All configurations are trained separately; this measures applicability across sizes, not zero-shot transfer. The direct Müller-style scalarization requires 6, 28, 120, and 496 scalar inputs for N=2,4,8,16, respectively, while lift + scalarization requires 10, 36, 136, and 528. In contrast, EQ-LNN learns interactions between representation-valued channels through its factorized bilinear layers, without constructing this complete set of scalar inputs in advance; the network stays compact (~75k–80k parameters across all tested SpringGraph sizes).

Larger, noisier, and more system-like settings

Noise To address the noiseless-training concern, we corrupted the TwoBodySpring training and validation observations with additive Gaussian noise and evaluated on a clean test pool:

Train noise sigmaEQ-LNN clean-test MAELifted LNN clean-test MAE
07.90e-54.14e-3
0.017.10e-31.49e-2
0.057.23e-29.44e-2

EQ-LNN maintains lower prediction error at both noise levels, although both models degrade substantially as noise increases. Importantly, the injected noise is large relative to the original prediction errors: even at σ=0.01\sigma=0.01, the positional and velocity perturbations have standard deviations of approximately 0.025 and 0.0067, compared with the clean EQ-LNN trajectory MAE of 7.90e-5. Furthermore, noise is independently applied to positions and velocities throughout the trajectories, introducing inconsistencies that a physically valid rollout cannot reproduce exactly. This creates a substantial noise floor in the training and validation objectives, making accurate recovery of the underlying dynamics more difficult. Thus, the observed degradation is understandable, although the experiment does not isolate this as its sole cause.

Despite this degradation, EQ-LNN retains the lower clean-test error, and its architectural conservation guarantee remains intact: the learned angular-momentum charge drifts by only ~1.6e-6, compared with 3.9–5.7 for Lifted LNN. We therefore distinguish predictive accuracy under noisy supervision from the exact structural properties of the learned model.

Spacecraft Our initial submission included several spacecraft systems in the appendix. SpacecraftA is a rigid bus with two hinged appendages (translation, MRP attitude, two hinge angles; 8 generalized coordinates). In the initial submission we noted that our full analytic lift succeeded on SpacecraftA while our minimal lift struggled. During the rebuttal we investigated this limitation and improved our minimal lift recipe by incorporating sparse marked-point kinematics.

Instead of the full analytic Cartesian geometry, the recipe asks for a handful of body-attached points: three shape vectors in the bus frame, plus a hinge offset and one marked point per appendage. A generic kinematic operator then composes these with the bus attitude and hinge rotations to get world-frame landmark positions and their velocities. This avoids the system-specific jacobian derivation, and only requires coarse geometry.

This lift has 45 features against 69 for the analytic lift, and no derived Jacobian. Even with fewer features, EQ-LNN gets back close to the analytic result: 4.14e-4 median trajectory MAE, versus 2.11e-2 for the matched Lifted LNN on identical features.

We thank the reviewer again for their constructive feedback. We hope the additional experiments and clarifications address the concerns.

As the discussion period draws to a close, we would like to thank the reviewer for their thoughtful feedback. We also wanted to highlight that we have added the direct invariant-scalar LNN comparison, as well as larger-system scaling, complex systems and noisy-data experiments. We also uploaded a revised manuscript with the mentioned experiments and updated clarifications and discussions around the additional experiments.

Since the comparison with Müller (2023) was your main concern, we would appreciate hearing whether the new experiments and clarification address it, or whether there are any remaining questions we could clarify during the discussion period. Thank you for your time and consideration.

Follow-up noise study:

After observing the results of our initial noise study, we wanted to explore further the effects of noise. We conducted two additional matched experiments on TwoBodySpring to complement the noisy-training results reported above.

First, we evaluated clean-trained models from initial states corrupted with 5% Gaussian observation noise. Comparing predictions against reference trajectories initialized from the same perturbed states, median EQ-LNN MAE increased from 7.90e-5 to 8.52e-5 versus 4.45e-3 for Lifted LNN under perturbed initialization. This was to test the case of training on clean data, but dynamics propagation from a noisy state (ie a setting useful for MPC or MBRL rollouts from noisy state)

Second, we wanted to test the affects of applying a standard smoothing filter to each algorithm. Recall, that at \sigma=0.05, EQ-LNN's clean-test MAE increased from 7.90e-5 to 0.0723 (Lifted LNN: 0.00414 to 0.0944). (As noted earlier, the observation noise exceeds the clean prediction error and makes dynamics learning more difficult.). Applying identical five-point Savitzky–Golay smoothing to the noisy training and validation data, (without modifying the clean test set) reduced these errors to 0.0311 for EQ-LNN and 0.0638 for Lifted LNN (five seeds). EQ-LNN achieved 2.05× lower error with the same preprocessing, while retaining near-exact learned angular-momentum conservation (normalized drift (1.48e-6)).

These experiments test robustness to noisy initial conditions at inference and learning accurate dynamics from noisy training observations. Clean-trained EQ-LNN remains highly accurate when tested from noisy rollout initialization its MAE increases by ~8% and remains 52× lower than Lifted LNN. Standard denoising also mitigates the degradation from noisy training, while EQ-LNN's architectural conservation guarantees remain intact.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback. As the discussion period comes to a close, we would like to summarize the updated experiments and revisions to the submission:

  • Improved SpacecraftA results: Updating the minimal lift operation for Eq-LNN achieves 4.14e-4 MAE vs. 2.11e-2 for lifted LNN
  • Scaling and topology: New experiments with 2–16 bodies and chain, ring, and star interactions
  • New baselines: Direct comparisons with Müller (2023) and GHNN (2025). We also developed an extension of the invariant-scalar LNN approach using our symmetry-compatible lifts, enabling comparisons on articulated and rigid-body systems beyond the original particle benchmarks.
  • Additional analyses: Noise robustness and conservation, computational cost, bilinear rank, and anchor-conditioned CoupledPendulum experiments

We also updated the related-work discussion, clarifited the lift methodology section, and expanded the description on spacecraft equations of motion.

We appreciate the reviewers taking the time to consider these additions and would be happy to address any remaining questions or concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.