← Paper Digest

Making Koopman Control Guarantees Hold on Robots

Abulikemu Abuduweili, Changliu Liu

November 2026

TL;DR A Koopman controller that is certified stable in its learned latent space can still be unstable on the real plant. Tangent-Invariant Koopman (TIK) forces the learned model to match a locally fitted linearization at the target. In four simulated benchmarks all 40 Standard Koopman controllers fail on the plant, and all 40 TIK controllers stabilize it.

What they do

Koopman control lifts nonlinear dynamics into a latent space where they are linear, so LQR and MPC become cheap. The stability check, however, is done on the latent closed loop, while the deployed controller re-encodes the measured state at every step. The authors show that these two loops can disagree. The latent certificate carries over to the plant only when the learned closed loop matches the plant’s closed-loop Jacobian on the tangent space of the embedding at the target.

How it works

  • Theory. With exact first-order matching, every eigenvalue of the plant’s closed loop also appears in the latent model. A margin covers approximate matching. A latent model that is open-loop stable cannot exactly match an unstable equilibrium.
  • TIK. A weighted least-squares fit on near-target offline transitions estimates the local plant linearization. The latent A and B are parameterized so that they match this fit exactly, and only the part off the tangent space is learned. TIK’s gain at the target therefore equals the LQR gain on the fit.
  • Large maneuvers. An optional bilinear term models state-dependent actuation. MPPI drives the system into a certified terminal set, where the controller switches to local LQR.

Results

  • All 40 Standard Koopman models are certified stable in latent space, yet every one is unstable on its plant. All 40 TIK models are stable.
  • Across 80 gain scalings, the latent and deployed verdicts agree for 14% of Standard Koopman model-gain pairs and 100% for TIK. TIK matches oracle LQR’s tilt-recovery outcome in all 420 tests.
  • Cart-pole swing-up succeeds 100% of the time with a 2 ms budget and 88% with 0.25 ms. MPPI-only swing-up succeeds 95% with the bilinear input and 6% with the linear one.
  • On simulated seven-joint arms TIK reaches 100% success at 1-5 ms per step. This matches linear MPC but does not beat it.

Limitations

Despite the title, every experiment runs in simulation; hardware noise and contact transitions are left for future work. The certificates need valid curvature and noise bounds, but the experiments use empirical estimates. The offline test is inconclusive under high noise. On the double cart-pole, TIK (83%) stays below fitted linear MPC (90%).

Figures

From Figure 1: First-pole angle of a double cart-pole after a 0.6 degree tilt, for Standard Koopman and TIK. Dashed l...
From Figure 1: First-pole angle of a double cart-pole after a 0.6 degree tilt, for Standard Koopman and TIK. Dashed lines are each controller's latent prediction and solid lines the simulated plant. The Standard Koopman plant trace ends when the pole falls.
From Figure 2: (a) Spectral radius predicted by the latent model against the deployed spectral radius at the nominal ...
From Figure 2: (a) Spectral radius predicted by the latent model against the deployed spectral radius at the nominal LQR gain, for Standard Koopman (orange), Koopman + invariance penalty (gray) and TIK (blue). (b) Percentage of models (40 seeds per method) whose latent stability prediction matches the true deployed stability as the feedback gain is scaled from 0.05 to 4.
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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.