- 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
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.
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.
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.