cs.ROFeb 23, 2026

Scaling Law of Neural Koopman Operators

Authors: Abulikemu Abuduweili, Yuyang Pang, Feihan Li, Changliu Liu

Organizations: Authors are with the Robotics Institute, Carnegie Mellon University, Pittsburgh, PA 15213, USA.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Data-driven neural Koopman operator theory has emerged as a powerful tool for linearizing and controlling nonlinear robotic systems. However, the performance of these data-driven models fundamentally depends on the trade-off between sample size and model dimensions, a relationship for which the scaling laws have remained unclear. This paper establishes a rigorous framework to address this challenge by deriving and empirically validating scaling laws that connect sample size, latent space dimension, and downstream control quality. We derive a theoretical upper bound on the Koopman approximation error, explicitly decomposing it into sampling error and projection error. We show that these terms decay at specific rates relative to dataset size and latent dimension, providing a rigorous basis for the scaling law. Based on the theoretical results, we introduce two lightweight regularizers for the neural Koopman operator: a covariance loss to help stabilize the learned latent features and an inverse control loss to ensure the model aligns with physical actuation. The results from systematic experiments across six robotic environments confirm that model fitting error follows the derived scaling laws, and the regularizers improve dynamic model fitting fidelity, with enhanced closed-loop control performance. Together, our results provide a simple recipe for allocating effort between data collection and model capacity when learning Koopman dynamics for control.

Figures & tables

Fig. 1 : Overview. (a) Mechanical systems (e.g., robot arm, quadruped) exhibiting complex nonlinear dynamics in physical state space.e. (b) Our proposed Koopman architecture uses an encoder network Ψθ\Psi_{\theta} to map the physical state xtx_{t} into a latent space ztz_{t} , governed by linear dynamics: zt+1=Azt+Butz_{t+1}=\mathbf{A}z_{t}+\mathbf{B}u_{t} .
EnvironmentState Dim.Action Dim.
Damping Pendulum21
Double Pendulum42
Franka Panda147
Kinova Gen3147
Unitree Go23512
Unitree G15323
TABLE I : Summary of environment dimensions. Note: Unitree G1 and Go2 train on trimmed state-action spaces.
Fig. 2 : Scaling of prediction error with training sample size ( mm ). Subplots display error curves across latent dimension multipliers ( nmultin_{\text{multi}} ). Dots denote random seeds, and solid lines show fitted power-law trends. The legend reports the scaling exponent α\alpha (error decay rate). Increasing dataset size consistently reduces error via a power law across environments, empirically validating theoretical sampling error convergence.
Fig. 3 : Scaling of prediction error with latent dimension nn . Subplots display error curves across training sample sizes ( mm ). Increasing latent dimension generally reduces error, empirically validating projection error convergence as Koopman subspace capacity expands.
Environmentαm\alpha_{m} (Sample Size)αn\alpha_{n} (Latent Dimension)
Damping Pendulum1.154.3
Double Pendulum0.752.89
Franka0.491.16
Kinova1.442.58
Go20.650.47
G10.520.34
TABLE II : Fitted power-law exponents ( α\alpha ) for prediction error decay. Sample efficiency αm\alpha_{m} is averaged across latent dimensions and seeds. Spectral decay αn\alpha_{n} is averaged across seeds in high-data regimes ( m≥104m\geq 10^{4} ), excluding smaller samples to isolate asymptotic projection error from finite-sample overfitting.
Fig. 4 : Scaling of prediction error with latent dimension ( nn ) under the coupled scaling law m=coeff⋅nln⁡nm=\text{coeff}\cdot n\ln n . As the data sufficiency coefficient increases, the error curves converge toward the asymptotic full sample baseline.
Fig. 5 : Franka arm tracking error versus training samples (left) and latent dimension (right). Auxiliary losses (+Both) yield consistent performance boosts.
Fig. 6 : G1 performance scales positively with sample size (left) and latent capacity (right). Combined auxiliary objectives (+Both) achieve the highest survival rates.
Fig. 7 : Go2 control performance (survival steps). Auxiliary losses enable better scaling, especially in high-capacity regimes (right) where the baseline struggles.
MetricsEnvironment+ Cov loss+ Ctrl loss+ Both
Track Err ( ↓\downarrow )Franka6.18%14.30%25.09%
Survival Rate ( ↑\uparrow )G12.09%7.29%9.60%
Go213.6%59.62%74.76%
TABLE III : Relative control performance improvement over baseline via Covariance ( Lcov\mathcal{L}_{\text{cov}} ) and Inverse Control ( Lctrl\mathcal{L}_{\text{ctrl}} ) auxiliary losses.
EnvironemntEDMDNNDMKoopmanKoopman
(Poly)(MLP)(Base NN)(+Aux)
Franka0.4410.2680.1220.121
G10.2710.1160.1300.132
Go20.4280.1630.2040.209
TABLE IV : Open-loop prediction error comparing classical EDMD (polynomial), unstructured Neural Dynamics (NNDM), and our Neural Koopman architecture.
MetricEnvEDMDNNDMKoopmanKoopman
(Poly)(MLP)(Base NN)(+Aux)
Track Err( ↓\downarrow )Franka0.4420.5140.4370.256
Survival( ↑\uparrow )G134.027.458.059.7
Go219.54.762.0103.0
TABLE V : Closed-loop control performance comparing tracking error (Franka) and survival steps (G1, Go2). Neural Koopman combines deep learning expressivity with Linear MPC.
Fig. 8 : Gram matrix condition number κ(G)\kappa(\mathbf{G}) versus prediction error. Marker size and opacity denote sample size mm (smaller/darker indicates larger mm ). Lines show fitted trends with (orange) and without (blue) covariance regularization. Grey arrows connect identical sample sizes. The proposed covariance loss significantly reduces κ(G)\kappa(\mathbf{G}) (leftward shift) and consistently correlates with lower error.
Fig. 9 : Control matrix condition number κ(B⊤B)\kappa(\mathbf{B}^{\top}\mathbf{B}) versus prediction error. Marker size/opacity denote sample size mm (smaller/darker indicates larger mm ), and grey arrows connect identical sample sizes. Lines show fitted trends with (orange) and without (blue) inverse control regularization. Unlike the Gram matrix, κ(B⊤B)\kappa(\mathbf{B}^{\top}\mathbf{B}) correlates negatively with error; better models naturally evolve higher condition numbers.
EnvironmentState Dim.Action Dim.
Polynomial30
Damping Pendulum21
Double Pendulum42
Franka Panda147
Kinova Gen3147
Unitree Go23512
TABLE VI : Summary of environments and original dimensions. Note that the Unitree G1 and Go2 models are trained on trimmed state–action spaces.
npolyn_{\text{poly}}Sample Size (αm,Cm,R2)(\alpha_{m},C_{m},R^{2})Latent Dim (αn,Cn,R2)(\alpha_{n},C_{n},R^{2})
3(2.64,  1.75×10−8,  1.00)(2.64,\;1.75{\times}10^{-8},\;1.00)(0.28,  1.18×10−21,  0.25)(0.28,\;1.18{\times}10^{-21},\;0.25)
5(2.57,  8.79×10−8,  1.00)(2.57,\;8.79{\times}10^{-8},\;1.00)(0.20,  2.72×10−21,  0.51)(0.20,\;2.72{\times}10^{-21},\;0.51)
10(2.52,  4.74×10−7,  1.00)(2.52,\;4.74{\times}10^{-7},\;1.00)(0.14,  4.51×10−20,  0.74)(0.14,\;4.51{\times}10^{-20},\;0.74)
20(1.91,  2.98×10−6,  1.00)(1.91,\;2.98{\times}10^{-6},\;1.00)(0.08,  4.54×10−20,  0.57)(0.08,\;4.54{\times}10^{-20},\;0.57)
50(1.64,  2.63×10−5,  0.99)(1.64,\;2.63{\times}10^{-5},\;0.99)(0.02,  4.17×10−19,  0.53)(0.02,\;4.17{\times}10^{-19},\;0.53)
TABLE VII : Impact of Nonlinearity on Scaling. Higher npolyn_{\text{poly}} leads to slower exponent decay term (lower α\alpha ) and higher error floors ( CC ). R2R^{2} indicates fit reliability.
Degree ( npolyn_{\text{poly}} )Baseline+ Cov lossError Reduction
39.51×10−69.51\times 10^{-6}8.89×10−68.89\times 10^{-6}6.5%6.5\%
52.12×10−52.12\times 10^{-5}2.04×10−52.04\times 10^{-5}4%4\%
106.22×10−46.22\times 10^{-4}5.47×10−45.47\times 10^{-4}12.1%12.1\%
201.34×10−21.34\times 10^{-2}1.17×10−21.17\times 10^{-2}12.7%12.7\%
501.98×10−11.98\times 10^{-1}1.62×10−11.62\times 10^{-1}18.1%18.1\%
TABLE VIII : Impact of covariance regularization loss ( Lcov\mathcal{L}_{\text{cov}} ) on prediction error. Evaluated across varying polynomial degrees ( npolyn_{\text{poly}} ) on largest model capacity ( nmulti=16n_{\text{multi}}=16 ) and dataset size ( m=140km=140\text{k} ).
Fig. 10 : Impact of Covariance Regularization on Latent Feature Correlation. Heatmaps display the correlation matrices of the learned embeddings ϕ(x)\phi(x) for varying polynomial degrees npoly∈{3,10,50}n_{\text{poly}}\in\{3,10,50\} . Top Row (Baseline): Without regularization, features exhibit high cross-correlation (dense off-diagonal blocks). Bottom Row (w/ Lcov\mathcal{L}_{\text{cov}} ): With regularization, the matrices become significantly sparser, demonstrating effective feature decorrelation. The mean off-diagonal correlation ∣corr‾∣|\overline{corr}| is reported for each case.
More in this topic

Koopman Operator Learning

Koopman operator learning builds finite-dimensional linear models of nonlinear dynamics by finding useful observable representations from data.

See all 79 papers

Explore similar work

CardsList
Apr 21, 2026

Efficient Reinforcement Learning using Linear Koopman Dynamics for Nonlinear Robotic Systems

This paper presents a model-based reinforcement learning (RL) framework for optimal closed-loop control of nonlinear robotic systems. The proposed approach learns linear lifted dynamics through Koopman operator theory and integrates the resulting model into an actor-critic architecture for policy optimization, where the policy represents a parameterized closed-loop controller. To reduce computational cost and mitigate model rollout errors, policy gradients are estimated using one-step predictions of the learned dynamics rather than multi-step propagation. This leads to an online mini-batch policy gradient framework that enables policy improvement from streamed interaction data. The proposed framework is evaluated on several simulated nonlinear control benchmarks and two real-world hardware platforms, including a Kinova Gen3 robotic arm and a Unitree Go1 quadruped. Experimental results demonstrate improved sample efficiency over model-free RL baselines, superior control performance relative to model-based RL baselines, and control performance comparable to classical model-based methods that rely on exact system dynamics.

Wenjian Hao, Yuxuan Fang, Zehui Lu +1

School of Aeronautics and Astronautics, Purdue University, West Lafayette, IN 47907, USA · independent researcher

Robotic ControlReinforcement LearningKoopman Operator LearningRobot Dynamics LearningActor-Critic MethodsOptimal Control+1
AskPDF
Mar 4, 2026

Whole-Body Safe Control of Robotic Systems with Koopman Neural Dynamics

Controlling robots with strongly nonlinear, high-dimensional dynamics remains challenging, as direct nonlinear optimization with safety constraints is often intractable in real time. The Koopman operator offers a way to represent nonlinear systems linearly in a lifted space, enabling the use of efficient linear control. We propose a data-driven framework that learns a Koopman embedding and operator from data, and integrates the resulting linear model with the Safe Set Algorithm (SSA). This allows the tracking and safety constraints to be solved in a single quadratic program (QP), ensuring feasibility and optimality without a separate safety filter. We validate the method on a Kinova Gen3 manipulator and a Go2 quadruped, showing accurate tracking and obstacle avoidance.

Sebin Jung, Abulikemu Abuduweili, Jiaxing Li +1

Robotics Institute, Carnegie Mellon University, Pittsburgh, PA.

Robotic ControlRobot SafetyWhole-Body ControlRobot Dynamics LearningKoopman Operator Learning
AskPDF
Aug 27, 2026

Data-driven Koopman mode approximation: A neural power iteration algorithm

This paper proposes a novel data-driven algorithm to approximate the dominant eigenfunctions (aka.~modes) of the Koopman operator of nonlinear dynamical systems using neural networks. The relevance of learning the dominant Koopman modes is to approximate nonlinear dynamics by linear ones in a lifted space, thereby enabling simplified control and analysis. To fight the curse of dimensionality arising from using expressive templates (here neural networks) for the mode approximation, the proposed method leverages a power-iteration scheme that directly learns the dominant Koopman modes without explicitly constructing the projection of the Koopman operator on the template of functions. Our approach connects to other approaches in the literature that avoid the curse of dimensionality by learning small dictionaries of functions, but differs from them in that we do not require ``anti-collapse mechanisms'' to ensure that the learned dictionary is expressive enough to approximate the Koopman operator since our power-iteration scheme is designed to converge toward the dominant modes of the projected Koopman operator. The approach is fully data-driven, requiring only sampled state transitions. Theoretical guarantees are provided, showing convergence under increasing sample size and network width (in connection with the neural tangent kernel theorem). Numerical experiments demonstrate that the method achieves accurate and smooth approximations of dominant modes while avoiding the limitations of traditional techniques such as extended dynamic mode decomposition.

Guillaume O. Berger, Raphaël M. Jungers

ICTEAM, UCLouvain, Louvain-la-Neuve, Belgium

Koopman Operator LearningSpectral Methods
AskPDF