Equivariant Graph Mechanics Networks with Constraints

1 INTRODUCTION

• Interacting systems also pose combinatorial complexity and require equivariance to translations, reflections, and rotations.GNN message passing addresses arbitrarily ordered objects and combinatorial relations.Source: §1 · p. 2
• Geometrically constrained systems are important in robotics and biochemistry, where broken joint or bond constraints can make predictions physically or chemically invalid.The paper identifies modeling these systems as an underexplored challenge for equivariant GNNs.Source: §1 · p. 2
• GMN uses generalized coordinates to encode structural-object constraints implicitly in forward kinematics, maintaining constraints without external regulation.A structural object is defined as multiple rigidly connected particles.Source: §1 · p. 2
• GMN learns interaction forces and generalized-coordinate accelerations with equivariant functions, whose universal approximation properties are theoretically discussed under specified conditions.The formulation is described as more computationally efficient than spherical-harmonic-based versions and more general than EGNN’s.Source: §1 · p. 2
• GMN is reported to outperform state-of-the-art models in prediction accuracy, constraint satisfaction, and data efficiency on simulated systems, with evaluations also on MD17 and CMU Motion Capture.The simulated dataset includes particles, sticks, and hinges.Source: §1 · p. 2

2 RELATED WORK

• Interaction Network, HRN, NRI, and Hamiltonian graph networks extend GNN-based physical simulation through interaction learning, hierarchical convolution, inferred interactions, and Hamiltonian mechanics.The passage identifies them as Battaglia et al. (2016), Mrowca et al. (2018), Kipf et al. (2018), and Sanchez-Gonzalez et al. (2019), respectively.Source: §2 · p. 2
• Prior symmetry-aware methods include translation-equivariant models, spherical-harmonic approaches such as Tensor-Field networks and SE(3) Transformer, Lie convolution, and EGNN.The passage notes that some methods enforce translation but not rotation equivariance; EGNN offers equivariant graph message passing without expensive higher-order representations.Source: §2 · p. 3
• Constraint-learning approaches use soft Lagrangian regulation, while DeLaN addresses one rigid object and DPI-Net’s box task uses canonical coordinates but scalar messages.The passage says these approaches do not study equivariant models; DeLaN does not target complex systems with multiple rigid and structural objects.Source: §2 · p. 3

3 GRAPH MECHANICS NEURAL NETWORK

• The method builds on particle states and graph interactions, with position and velocity treated as geometric vectors subject to translation and rotation equivariance.Each particle also has a non-vector feature channel, distinct from the kinematics vectors.Source: §3 · p. 3

3.1 OUR GENERAL ARCHITECTURE

• GMN updates structural objects through generalized coordinates, so connected particles move jointly and geometric constraints are represented through their kinematics.It computes particle interactions, aggregates information within each object to infer generalized acceleration, updates those coordinates, and maps them back to particle states by forward kinematics.Source: §3.1 · p. 4

3.2 EQUIVARIANT MESSAGE PASSING

• The general equivariant form extends beyond a single input vector, addressing a limitation left open by EGNN's particular acceleration formulation.The paper also reports that adding a normalization term when m > 1 gives more stable performance without changing the stated theoretical conclusions.Source: §3.2 · pp. 5–6
• The proposed orthogonality-equivariant function is universally expressive when the input vectors span the full space; with fewer vectors, this guarantee applies within their span.The authors note that EGNN's single-relative-position form is not universally expressive, although it remains physically complete for Coulomb forces oriented by that position.Source: §3.2 · pp. 5–6

3.3 IMPLEMENTATIONS OF FORWARD KINEMATICS

• Hinges use a free Cartesian coordinate for one particle and angle coordinates for the other two, while sticks use a center coordinate and shared rotation angles.For hinges, angular acceleration is computed from relative torque and moment of inertia; the stick center can use a simplified acceleration based on the forces.Source: §3.3 · p. 6
• A learnable forward-kinematics variant replaces the hand-crafted updates with an equivariant function.The supplied passage identifies the variant but refers its full details to Appendix I.Source: §3.3 · p. 6

4.1 SIMULATION DATASET: CONSTRAINED N-BODY

• The simulated task predicts particle positions at T = 1000 across systems with 5, 10, or 20 particles, sticks, hinges, and training sets of 500 or 1500 trajectories.The evaluation compares models across object combinations and training-set sizes.Source: §4.1 · p. 7

4.1 SIMULATION DATASET: CONSTRAINED N-BODY; 4.2 APPLICATIONS ON REAL-WORLD DATASETS

• GMN is best across simulated scenarios, outperforms other competitive equivariant models on 7 of 8 MD17 molecules, and beats other models by a large margin on motion capture.The real-world evaluations report prediction errors averaged across three runs.Source: §4.1 · p. 7, §4.2 · p. 9, Table 4 · p. 9

4.1 SIMULATION DATASET: CONSTRAINED N-BODY

• On (3,2,1), as training size varies, GMN acts steadily; EGNN and EGNNReg perform much worse with small training sets, then approach GMN when the dataset is enlarged sufficiently.Source: §4.1 · p. 8
• When trained on (3,2,1) and tested on other systems, GMN's transfer performance is close to its original-setting performance, and GMN performs best.The tested systems are (2,4,0) and (1,0,3).Source: §4.1 · p. 8, Table 2 · p. 8
• Removing equivariance from ϕ2 harms performance, and removing the normalization term causes training divergence; GMN-L still slightly trails GMN.GMN-L replaces hand-crafted forward kinematics with a learnable equivariant function and outperforms EGNN in various settings.Source: §4.1 · p. 8, Table 3 · p. 8

4.2 APPLICATIONS ON REAL-WORLD DATASETS

• The real-world application represents systems as particles and disjoint sticks for partial length constraints, avoiding manual derivation of full kinematics.The approach is applied to MD17 and CMU Motion Capture without modifying GMN's current formulation.Source: §4.2 · p. 8
• For MD17, GMN performs worse than other models on Ethanol, while GMN-L surpasses GMN on 4 of 8 molecules.The authors suggest bond constraints may hinder learning for small, simple Ethanol, but report more significant GMN improvements on complex molecules including Aspirin, Benzene, and Salicylic.Source: §4.2 · p. 9, Table 4 · p. 9
• For human motion, the study uses walking trials from subject #35 and treats body parts as hard rigid-body constraints; GMN-L is inferior to GMN but markedly better than other methods.GMN predictions are described as accurate in the visualization, while EGNN has larger error.Source: §4.2 · p. 9, Table 5 · p. 9, Figure 6 · p. 9

5 CONCLUSION

• Generalized coordinates encode geometrical constraints implicitly and exactly in forward kinematics.The representation describes positions and velocities of structural objects.Source: §5 · p. 9
• GMN develops orthogonality-equivariant functions for interaction forces and backward dynamics, with theoretical justification for their expressivity.Source: §5 · p. 9
• On simulated particles, sticks, and hinges, GMN outperforms existing methods in prediction error, constraint satisfaction, and data efficiency; real-world evaluations support generalization to complex systems.Source: §5 · p. 9

7 ETHICS STATEMENT

• The research does not involve any human subject.Source: §7 · p. 10
• The authors state that their dataset is not related to privacy issues and can be used publicly.Source: §7 · p. 10
• All authors state that they follow the ICLR Code of Ethics.Source: §7 · p. 10

A THE PROOF OF THEOREM 1

• When m ≥ d and Z has full row rank, an MLP can approximate any continuous orthogonality-equivariant function ϕ̂(Z, h).The theorem states this conditional universality result for the proposed equivariant form.Source: A THE PROOF OF THEOREM 1 · p. 12
• The proof represents an equivariant output as Zπ(Z) and expresses invariant dependence through Z⊤Z before applying an MLP approximation.The proof invokes MLP universality to establish the approximation conclusion.Source: A THE PROOF OF THEOREM 1 · pp. 12–13
• If m ≥ d and Z's entries are independently drawn from an absolutely continuous distribution, the theorem's conclusion holds almost surely.The corollary relies on full rank occurring with probability one under these conditions.Source: A THE PROOF OF THEOREM 1 · p. 13
• The universality conclusion also applies to continuous orthogonality-equivariant functions whose outputs lie in the subspace spanned by Z's columns.Source: A THE PROOF OF THEOREM 1 · p. 13

B THE DYNAMICS ANALYSES FOR STICKS AND HINGES

• For sticks, center-of-mass acceleration follows the applied forces, while rotation acceleration is expressed using total torque and moment of inertia; hinge dynamics use the forces on three particles and their kinematic relations.The derivations assume equal particle masses and massless sticks.Source: B THE DYNAMICS ANALYSES FOR STICKS AND… · pp. 13–14

C KINEMATICS DECOMPOSITION

• Kinematics decomposition represents an arbitrary graph as particles and non-intersecting sticks, extending the formulation to complex systems while preserving only partial constraints.For CMU Motion Capture, the tree-like graph also permits exact forward kinematics through recursion from the root.Source: C KINEMATICS DECOMPOSITION · p. 14

D FULL ALGORITHMIC DETAILS

• Algorithm 1 combines the general GMN pipeline with implementation details, taking initial particle states and learnable equivariant functions through to predicted particle states.Its listed stages include computing generalized coordinates, interaction forces, and generalized accelerations before forward kinematics.Source: D FULL ALGORITHMIC DETAILS · p. 15, C KINEMATICS DECOMPOSITION · p. 15

• Most models use four layers, batch size 200, learning rate 0.0005, hidden dimension 64, and weight decay 1 × 10−10; SE(3)-Transformer and TFN use a separately tuned configuration.EGNNReg adds constraint error to its training loss, with λ selected from [0.01, 0.1].Source: E MORE EXPERIMENTAL DETAILS AND RESULTS · pp. 15–16
• For MD17, evaluation uses 500/2,000/2,000 train/validation/test frame pairs, masks hydrogen atoms, and selects non-intersecting bonds as sticks, leaving some bond lengths unpreserved.Inputs are separated from prediction frames by T = 5000; the setup augments molecular graphs with 2-hop neighbors.Source: E MORE EXPERIMENTAL DETAILS AND RESULTS · p. 16
• For CMU Motion Capture, the setup samples 200/600/600 train/validation/test frame pairs from 11/6/6 trials and models six key bones as sticks alongside 19 isolated particles.The frame-pair span is T = 30.Source: E MORE EXPERIMENTAL DETAILS AND RESULTS · p. 16
• In the simulation visualizations, GMN tracks ground-truth trajectories while EGNN shows position errors and constraint breaks; on motion capture, GMN predictions are more accurate than EGNN’s.The comparisons are shown in Figures 8 and 9.Source: E MORE EXPERIMENTAL DETAILS AND RESULTS · pp. 16–17, Figure 8 · p. 16, Figure 9 · p. 17
• GMN outperforms other models across all object combinations in Table 6, and has lower training and testing loss than EGNN and EGNNReg on the (3,2,1) learning curves.Table 6 reports prediction error for systems with different counts of particles, sticks, and hinges, trained with 500 samples.Source: Table 6 · p. 17, E MORE EXPERIMENTAL DETAILS AND RESULTS · p. 17, Figure 10 · p. 17

F MORE ABLATIONS

• Treating a hinge as two sticks performs worse than hinge-modeled GMN but better than EGNN, while assigning charges to nodes instead of edges makes very limited performance difference.The hinge split overlooks kinematics at the connected point; models can learn charge interactions from node features.Source: F MORE ABLATIONS · p. 18, Table 7 · p. 18, Table 8 · p. 18

G CONSTRAINT SATISFACTION

• GMN’s constraint errors are all below 1e-4; EGNNReg reduces errors, but its soft, data-driven constraints remain limited by training-sample counts.The errors measure average trajectory-level changes in stick and hinge lengths.Source: G CONSTRAINT SATISFACTION · p. 18, Table 9 · p. 18

H MORE DISCUSSIONS ON GENERALIZATION

• GMN transfer performance is very close to its original-setting performance when trained on (3,2,1) and tested across different scenarios.Source: H MORE DISCUSSIONS ON GENERALIZATION · p. 19, Table 10 · p. 19

• GMN-L replaces hand-crafted forward kinematics and Euler-angle computations with a parameterized equivariant design, while GMN-LReg adds explicit constraint regularization.The parameterized design uses object-level generalized coordinates and does not preserve constraints as exactly as exact FK.Source: I LEARNABLE FK · p. 19
• GMN-L consistently outperforms EGNN and its regularized version across settings, while GMN-L and GMN-LReg have a minor performance gap from GMN with exact FK.The authors interpret the gap as evidence that hard-coding constraints replaces a nontrivial amount of learning complexity.Source: I LEARNABLE FK · p. 19, Table 11 · p. 19

J ROBUSTNESS TO ERRORS IN THE PHYSICAL PRIOR OF CONSTRAINTS

• The robustness tests perturb stick connectivity by joining isolated particles, splitting sticks, doing both, or adding Gaussian noise to stick lengths.Length noise is N(0, 0.1L), where L is the original stick length; perturbations are independently applied for each training sample each time it enters the network.Source: J ROBUSTNESS TO ERRORS IN THE PHYSICAL… · pp. 19–20
• MSE differences ≤0.30 under the tested perturbations, which generally did not jeopardize GMN’s performance on noisy stick-constraint priors.The authors interpret this as insensitivity to slight errors in the input physical prior and an ability to learn to some degree with wrong constraints.Source: J ROBUSTNESS TO ERRORS IN THE PHYSICAL… · p. 20, Table 12 · p. 20

K MORE DISCUSSIONS ON THE DECOMPOSITION

• On MD17, GMN-LH is slightly worse than GMN-L on several molecules but gives desirable results on Ethanol and Benzene; on Motion Capture, it modestly beats GMN-L but remains below GMN.The results are shown in the MD17 and Motion Capture prediction-error tables, respectively.Source: K MORE DISCUSSIONS ON THE DECOMPOSITION · p. 20, Table 13 · p. 20, Table 14 · p. 20
• The authors recommend stick-based decomposition by default because sticks are the basic building blocks of hinges and other larger rigid objects.GMN-LH is the hinge-wise alternative, decomposing systems into particles and hinges.Source: K MORE DISCUSSIONS ON THE DECOMPOSITION · p. 20