Making Koopman Control Guarantees Hold on Robots

Abulikemu Abuduweili, Changliu LiuarXiv 2026 · arXiv

TL;DR

Koopman control often fails on real robots because a stable "latent" model doesn't guarantee a stable physical system. This paper introduces Tangent-Invariant Koopman (TIK), which enforces first-order derivative matching between the learned model and the physical plant at the target equilibrium, ensuring that theoretical stability guarantees actually hold during deployment.

What Problem Does This Paper Address?

Koopman operator theory is a popular way to control nonlinear robots by "lifting" their dynamics into a higher-dimensional latent space where the physics appear linear. This allows researchers to use efficient, well-understood tools like Model Predictive Control (MPC) and Linear Quadratic Regulators (LQR).

However, there is a dangerous "stability gap." A Koopman model might look perfectly stable in simulation (the latent loop), but when you put it on a robot (the deployed loop), the robot falls over. This happens because the latent model predicts its own future states, while the deployed controller constantly "re-encodes" the robot's actual physical state. If the derivatives of the learned model don't match the derivatives of the real physics at the target point, the act of re-encoding introduces errors that can drive the system to instability.

Methods

The authors propose Tangent-Invariant Koopman (TIK) to bridge this gap. The key technical contribution is ensuring Tangent Invariance: the first derivatives of the prediction error must vanish at the target equilibrium.

  1. Local Plant Fitting: Instead of just learning a global model, TIK first performs a local weighted least-squares fit of the plant's Jacobians ($A_f, B_f$) using offline transition data near the target.
  2. Derivative Matching: TIK uses a specific parameterization that forces the learned Koopman matrices ($A, B$) to match the local plant fit on the tangent space of the learned embedding.
  3. Bilinear Prediction: To handle large maneuvers (like a cart-pole swing-up) where the "effectiveness" of an action changes based on the state, TIK incorporates a bilinear term. This allows the model to remain linear in control while capturing state-dependent actuation.
  4. Stability Certificates: The paper derives a robustness margin based on the "real stability radius." It provides an offline test that uses empirical curvature estimates to verify if a controller will be locally stable before it is ever deployed.
  5. Certified Terminal Sets: They define an invariant region (an ellipsoid) where, once the robot enters, it is guaranteed to stay within constraints and converge to the target under local feedback.

Datasets

The researchers evaluated TIK across several simulated robotic benchmarks of increasing complexity:

  • Classic Unstable Benchmarks: Cart-pole (both ODE and MuJoCo versions), Double Cart-pole, and Double Pendulum.
  • High-Dimensional Manipulators: A simulated 7-joint Kinova Gen3 arm performing obstacle avoidance and torque-controlled reaching tasks.
  • Legged Locomotion: A Unitree Go2 quadruped (35D state, 12D input) tested for posture recovery under pushes and payload shifts.

Key Insights and Ideas

  • The "Slope" Matters More than the "Point": Low prediction error is not enough for stability. A model can be very accurate at predicting the next state but have the wrong "slope" (Jacobian), leading to a controller that over-corrects or under-corrects, eventually causing a crash.
  • The Stability Obstruction: A crucial theoretical finding (Corollary 3) is that an open-loop stable latent model cannot exactly match an open-loop unstable physical equilibrium. This explains why many "stable-by-design" Koopman variants fail to stabilize systems like upright pendulums.
  • Separation of Concerns: TIK effectively separates the problem of local stabilization (handled by the tangent matching) from global nonlinear prediction (handled by the learned observables and bilinear terms).

Key Results

  • Local Stability: In tests across 40 different seeds for unstable benchmarks, Standard Koopman failed 100% of the time to stabilize the physical plant (despite being stable in the latent space). TIK succeeded 100% of the time.
  • Compute Efficiency: TIK achieved an 88% success rate on the constrained cart-pole swing-up task with a tiny 0.25 ms planning budget. This outperforms neural NMPC methods which often exceed such tight time limits.
  • Complex Systems: On the 7-joint manipulator, TIK reached 100% success in obstacle avoidance with a median solve time of 2.1 ms.
  • Robustness: On the Unitree Go2 quadruped, TIK maintained a 67%–74% success rate for posture recovery even under sensor noise, friction changes, and 20% payload increases, significantly outperforming standard Koopman and Neural MPPI baselines.

Limitations and Future Work

  • Formal vs. Empirical Bounds: The stability certificates rely on knowing the maximum "curvature" (Hessian) of the dynamics. In this paper, these were estimated empirically; finding formal, non-conservative bounds for complex neural network models remains a challenge.
  • Nonsmooth Dynamics: The theory assumes the dynamics are differentiable. Real-world robots often involve "hard" contacts (like feet hitting the ground), which are nonsmooth. The authors noted that contact transitions in the quadruped experiments could violate differentiability assumptions.
  • Hardware Noise: While simulated noise was tested, the specific characteristics of real-world sensor noise and latency on physical hardware require further study to see how they affect the robustness margins.

Summary updated

Abstract

A Koopman controller can stabilize its learned latent-space linear model yet fail to stabilize the nonlinear plant. The learned model propagates predictions in latent space, while the deployed controller repeatedly re-encodes the measured physical state. We show that nominal closed-loop stability transfers to the nonlinear plant when the learned and true closed-loop dynamics agree to first order on the embedding's tangent space at the target. We derive a robustness margin for approximate matching and show that a stable open-loop latent model cannot exactly match an unstable equilibrium. We introduce Tangent-Invariant Koopman (TIK), which estimates a local plant linearization from offline transitions and enforces first-order matching to this estimate. TIK recovers the fitted LQR gain at the target, while a bilinear predictor captures state-dependent actuation during large maneuvers. Across four simulated unstable benchmarks, all 40 Standard Koopman controllers have stable latent closed loops, but their deployed nonlinear systems are locally unstable. In contrast, all 40 TIK controllers locally stabilize their respective nonlinear plants. An offline stability test using empirical curvature estimates identifies stable TIK controllers on two cart-pole systems, but is inconclusive under high noise. Under explicit model-error and curvature bounds, an invariant terminal set guarantees constraint satisfaction after entry under local feedback. TIK achieves 88\% constrained cart-pole swing-up success with a 0.25 ms planning budget and controls simulated seven-joint arms within 1-5 ms per step.
BibTeX
@misc{abuduweili2026making,
  author = {Abulikemu Abuduweili and Changliu Liu},
  title = {{Making Koopman Control Guarantees Hold on Robots}},
  year = {2026},
  eprint = {2611.20402},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20402},
}
Figure