Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning

Christian Ritter; Jan Peters; Michael Lutter

arxiv: 1907.04490 · v1 · pith:5TJTONPNnew · submitted 2019-07-10 · 💻 cs.LG · cs.RO· cs.SY· eess.SY· stat.ML

Deep Lagrangian Networks: Using Physics as Model Prior for Deep Learning

Michael Lutter , Christian Ritter , Jan Peters This is my paper

Pith reviewed 2026-05-24 23:40 UTC · model grok-4.3

classification 💻 cs.LG cs.ROcs.SYeess.SYstat.ML

keywords deep learningLagrangian mechanicsdynamics learningrobot controlphysics priorsmodel-based controlsystem identification

0 comments

The pith

Structuring deep networks around Lagrangian mechanics allows them to learn mechanical system dynamics efficiently while preserving physical consistency.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes Deep Lagrangian Networks as a way to embed the structure of Lagrangian mechanics directly into a neural network for learning the dynamics of mechanical systems. By having the network output the kinetic energy, potential energy, and derived quantities like the inertia matrix, the approach ensures that learned models obey physical laws such as energy conservation and symmetric inertia. This structure supports learning from fewer samples, more reliable prediction on unseen trajectories, and real-time adaptation, which matters for applications like robot control where errors can be costly and data is often limited or collected online.

Core claim

Deep Lagrangian Networks (DeLaN) impose the Lagrangian formulation L = T - V on a deep network architecture so that it outputs the inertia matrix, Coriolis forces, and gravity terms in a form that automatically satisfies the properties required by Lagrangian mechanics. The resulting model can therefore learn the equations of motion of a mechanical system from data while guaranteeing physical plausibility, leading to faster learning, improved extrapolation, and the ability to update the model online during robot operation.

What carries the argument

Deep Lagrangian Networks (DeLaN), a neural network structured to compute the Lagrangian energy terms T and V and derive the equations of motion while enforcing physical constraints such as inertia matrix symmetry.

If this is right

The learned dynamics models support high-performance tracking control on robotic systems.
Fewer training samples are needed compared with unstructured deep networks for the same accuracy.
Extrapolation to trajectories outside the training distribution remains accurate and stable.
The model can be updated online in real time without losing physical consistency.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same structuring principle could be applied to other domains where conservation laws or symmetries are known, such as rigid-body dynamics in different coordinate systems.
It may reduce the data volume required for sim-to-real transfer in control tasks by baking in domain knowledge upfront.
Combining DeLaN with other priors, like contact constraints, could further improve robustness in complex environments.

Load-bearing premise

The mechanical system must obey Lagrangian mechanics, and the network parameterization must produce valid energy functions that yield consistent dynamics without additional violations.

What would settle it

A DeLaN model trained on trajectories from a simple mechanical system like a pendulum that then produces predictions violating energy conservation or symmetry of the inertia matrix on held-out data would falsify the physical-plausibility claim.

read the original abstract

Deep learning has achieved astonishing results on many tasks with large amounts of data and generalization within the proximity of training data. For many important real-world applications, these requirements are unfeasible and additional prior knowledge on the task domain is required to overcome the resulting problems. In particular, learning physics models for model-based control requires robust extrapolation from fewer samples - often collected online in real-time - and model errors may lead to drastic damages of the system. Directly incorporating physical insight has enabled us to obtain a novel deep model learning approach that extrapolates well while requiring fewer samples. As a first example, we propose Deep Lagrangian Networks (DeLaN) as a deep network structure upon which Lagrangian Mechanics have been imposed. DeLaN can learn the equations of motion of a mechanical system (i.e., system dynamics) with a deep network efficiently while ensuring physical plausibility. The resulting DeLaN network performs very well at robot tracking control. The proposed method did not only outperform previous model learning approaches at learning speed but exhibits substantially improved and more robust extrapolation to novel trajectories and learns online in real-time

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DeLaN bakes Lagrangian structure into the network so dynamics come out with the right symmetries and positive-definiteness by construction.

read the letter

The core move is to split the network into two heads that predict kinetic energy T(q, q̇) and potential V(q), then analytically compute M(q), C(q, q̇), and g(q) from the Euler-Lagrange equation. That guarantees the learned model obeys the structural properties of Lagrangian mechanics without extra loss terms or post-processing. The paper shows this on robot tracking tasks and reports faster learning plus better extrapolation to unseen trajectories than standard black-box models. It also claims real-time online adaptation works in practice. That combination of built-in physics and empirical control performance is the main contribution worth noting. The construction is cleaner than adding physics penalties to a generic network, and the guarantees on symmetry and skew-symmetry are automatic once the inertia matrix stays positive definite. The abstract and stress-test both indicate the derivation holds without internal contradictions. The main soft spot is that the provided abstract gives almost no experimental detail: no table of baselines, no error bars, no description of how positive-definiteness is enforced in the network, and no ablation on whether the analytic derivation is what drives the extrapolation gains. Without those sections it is difficult to judge how much the structure actually buys versus careful tuning or dataset choice. The claims about outperforming prior model-learning methods therefore rest on evidence that is not visible here. This work is aimed at people doing model-based control or system identification in robotics who already care about physical consistency. A reader who wants a dynamics model that cannot violate basic mechanical properties will find the architecture useful even if the experiments need tightening. The idea is grounded enough and the guarantees are real enough that it should go to peer review rather than a desk reject.

Referee Report

2 major / 2 minor

Summary. The paper introduces Deep Lagrangian Networks (DeLaN), a deep network architecture that embeds Lagrangian mechanics as a structural prior for learning system dynamics. Separate networks model the kinetic energy T(q, q̇) and potential energy V(q); the inertia matrix M(q), Coriolis matrix C(q, q̇), and gravity vector g(q) are then obtained analytically via the Euler-Lagrange equation. This construction guarantees that learned models satisfy key physical properties (symmetric positive-definite M, skew-symmetry of Ṁ−2C) by design. The method is applied to robot tracking control, with claims of faster learning, substantially improved extrapolation to novel trajectories, and real-time online adaptation compared to prior model-learning baselines.

Significance. If the empirical claims hold, the work is significant because it shows how an external physics prior can be imposed on a deep network to obtain data-efficient, physically consistent dynamics models that extrapolate reliably—addressing a central weakness of black-box deep learning in model-based control. The analytic derivation of the equations of motion from energy networks is a clean way to enforce structural invariants without post-hoc projection.

major comments (2)

[architecture / network structure] The manuscript states that the networks are structured to output valid energies yielding a positive-definite M(q), but the precise parameterization (e.g., Cholesky factorization or soft-plus on eigenvalues) and any associated regularization are not detailed in the architecture description; without this, it is unclear whether positive-definiteness is strictly enforced or only encouraged during training.
[experimental evaluation] The extrapolation and online-learning claims rest on the assumption that the true system exactly obeys Lagrangian mechanics with configuration-dependent inertia; if unmodeled effects (friction, actuator dynamics, or non-holonomic constraints) are present, the structural guarantee may not translate to the reported robustness gains. The experiments should quantify the magnitude of such violations.

minor comments (2)

[throughout] Notation for the learned energies (T̂, V̂) versus the derived quantities (M̂, Ĉ, ĝ) should be introduced once and used consistently; occasional reuse of symbols for network outputs and derived matrices reduces readability.
[experiments] The abstract claims “outperformed previous model learning approaches at learning speed,” yet the main text should explicitly list the baselines, their hyper-parameter budgets, and the precise metric (wall-clock time to reach a given tracking error) used for this comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and the recommendation of minor revision. The comments are constructive and we address each one below.

read point-by-point responses

Referee: [architecture / network structure] The manuscript states that the networks are structured to output valid energies yielding a positive-definite M(q), but the precise parameterization (e.g., Cholesky factorization or soft-plus on eigenvalues) and any associated regularization are not detailed in the architecture description; without this, it is unclear whether positive-definiteness is strictly enforced or only encouraged during training.

Authors: We agree that the architecture description would benefit from greater explicitness. In the revised manuscript we will expand Section 3.2 to state that the kinetic-energy network directly parametrizes the lower-triangular Cholesky factor L(q) of M(q), a softplus activation is applied to the diagonal entries of L(q) to guarantee strict positive-definiteness, and no auxiliary regularization beyond the standard supervised loss is used. This makes the enforcement of positive-definiteness a structural property rather than a training-time encouragement. revision: yes
Referee: [experimental evaluation] The extrapolation and online-learning claims rest on the assumption that the true system exactly obeys Lagrangian mechanics with configuration-dependent inertia; if unmodeled effects (friction, actuator dynamics, or non-holonomic constraints) are present, the structural guarantee may not translate to the reported robustness gains. The experiments should quantify the magnitude of such violations.

Authors: The referee correctly identifies that the structural guarantees presuppose Lagrangian dynamics. All simulated experiments satisfy this assumption exactly; the real-robot experiments include typical unmodeled effects (joint friction, actuator dynamics) yet still exhibit the reported gains in extrapolation and online adaptation. We will add a dedicated limitations paragraph acknowledging that the magnitude of these violations was not separately quantified. Performing such a quantification would require additional controlled experiments outside the scope of the present study; the comparative results nevertheless demonstrate that the physics prior remains beneficial even when the assumption is only approximately satisfied. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central construction defines separate networks to output kinetic energy T(q, q̇) and potential energy V(q), then analytically derives the inertia matrix M(q), Coriolis C(q, q̇) and gravity g(q) via the Euler-Lagrange equation. This imposes the external Lagrangian prior L = T - V by architecture, guaranteeing structural properties (symmetric positive-definite M, skew-symmetry of Ṁ - 2C) by construction from an independent physics domain rather than from data fits or self-citations. No step reduces a claimed prediction to a fitted input by definition, and no load-bearing self-citation or ansatz smuggling is present. The derivation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the standard domain assumption of Lagrangian mechanics for mechanical systems and the capacity of a neural network to represent the required energy functions under that structure. No new entities are postulated.

free parameters (1)

Neural network parameters
Weights and biases of the deep network that are optimized during training to match observed trajectories while respecting the imposed Lagrangian structure.

axioms (1)

domain assumption Mechanical systems follow Lagrangian mechanics, with equations of motion derived from L = T - V
This is the core prior directly imposed on the network structure as stated in the abstract.

pith-pipeline@v0.9.0 · 5732 in / 1369 out tokens · 35823 ms · 2026-05-24T23:40:50.412307+00:00 · methodology

discussion (0)

Forward citations

Cited by 15 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Embedding Hybrid Systems into Continuous Latent Vector Fields
cs.LG 2026-06 unverdicted novelty 7.0

An n-dimensional hybrid system embeds into a continuous vector field in m > 2n dimensions, enabling latent Neural ODEs with consistency losses to recover hybrid flows from time series.
Support-Safe Variational Hybrid Filtering for Contact-Mode and Sparse-Law Recovery
cs.RO 2026-05 unverdicted novelty 7.0

VHYDRO is a support-safe variational hybrid filter that jointly recovers continuous latent states, discrete contact modes, and sparse port-Hamiltonian laws per regime while preventing loss of feasible transitions.
Detecting Deepfakes via Hamiltonian Dynamics
cs.CV 2026-05 unverdicted novelty 7.0

HAAD detects deepfakes by modeling latent manifolds as potential energy surfaces and quantifying instability via Hamiltonian trajectory statistics such as action and energy dissipation.
RigPI: Dynamic Parameter Identification of Rigid Body via VLM-Seeded Differentiable Simulation
cs.RO 2026-06 unverdicted novelty 6.0

RigPI combines VLM semantic priors with two-stage gradient optimization in differentiable simulation to identify inertial and frictional parameters of rigid bodies from robot-object interactions.
NeuROK: Generative 4D Neural Object Kinematics
cs.CV 2026-05 unverdicted novelty 6.0

NeuROK learns a data-driven latent kinematic parameterization on a large 4D dataset to generate realistic object deformations by simulating dynamics only in low-dimensional latent space via Lagrangian mechanics.
Neural Co-state Policies: Structuring Hidden States in Recurrent Reinforcement Learning
cs.LG 2026-05 unverdicted novelty 6.0

Hidden states in recurrent RL policies correspond to PMP co-states, so a derived co-state loss structures the dynamics and yields robust performance on partially observable continuous control tasks.
QuietWalk: Physics-Informed Reinforcement Learning for Ground Reaction Force-Aware Humanoid Locomotion Under Diverse Footwear
cs.RO 2026-04 unverdicted novelty 6.0

QuietWalk combines an inverse-dynamics-constrained PINN for GRF estimation with RL to produce low-impact humanoid locomotion policies that generalize across footwear, cutting mean noise by 7.17 dB on hardware.
Dissipative Latent Residual Physics-Informed Neural Networks for Modeling and Identification of Electromechanical Systems
cs.LG 2026-04 unverdicted novelty 6.0

DiLaR-PINN learns dissipative effects in electromechanical systems via a skew-dissipative latent residual PINN that guarantees non-increasing energy and uses recurrent curriculum training for partial observations.
Sensorless Four-Channel Control Architecture Using Inverse Dynamics Modeling for Human-Scale Bilateral Teleoperation
cs.RO 2026-07 unverdicted novelty 5.0

A sensorless four-channel bilateral teleoperation architecture based on inverse dynamics modeling outperforms standard two- and four-channel methods in tracking, operator effort, and transmittable impedance on a human...
RigPI: Dynamic Parameter Identification of Rigid Body via VLM-Seeded Differentiable Simulation
cs.RO 2026-06 unverdicted novelty 5.0

RigPI combines VLM initialization with two-stage gradient-based optimization in differentiable simulation to estimate dynamic parameters of rigid bodies from real robot interactions.
Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems
cs.LG 2026-06 unverdicted novelty 5.0

GraMO couples graph interactions and temporal state updates in one linear recurrence with input-dependent coefficients to simulate N-body, motion, and robotics systems with lower long-horizon error than prior GNN or S...
Robots Need More than VLA and World Models
cs.RO 2026-06 unverdicted novelty 5.0

The paper identifies four missing interfaces (data autolabelling, embodiment retargeting, physics-grounded world models, and video-based reward inference) as the central bottleneck beyond VLA scaling for robot intelligence.
Neural Co-state Policies: Structuring Hidden States in Recurrent Reinforcement Learning
cs.LG 2026-05 unverdicted novelty 5.0

Recurrent RL policies can have their hidden states aligned with PMP co-states through a derived loss, yielding robust performance on partially observable control tasks.
Optimal transport by a Lagrangian dynamics of population distribution
cond-mat.dis-nn 2025-10 unverdicted novelty 5.0

A quadratic Lagrangian incorporating dissipation models human mobility from population distributions and fits both synthetic and empirical data, showing comparable inertia and dissipation effects.
World Models: A Comprehensive Survey of Architectures, Methodologies, Reasoning Paradigms, and Applications
cs.LG 2026-05 unverdicted novelty 3.0

The paper delivers a multi-axis taxonomy for world models that maps architectures, training families, reasoning strategies, and domains from early cognitive foundations through systems such as Dreamer, MuZero, and Sor...