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.
Figures & tables
Figure 1: The nominal and deployed loops can disagree. (a) The learned model propagates its own predictions. (b) Deployment feeds back the re-encoded plant state. (c) First-pole angle for Standard Koopman and TIK on shared axes (seed 42 , initial tilt 0.6∘ ). Each controller has a latent prediction (dashed) and a simulated plant rollout (solid). The Standard Koopman plant trace ends when the pole falls. (d) The true linearized image of a tangent vector lies in range(Jθ) . A learned update can leave this space. TIK sets AJθv=JθA^fv , which equals the true tangent action when the fit is exact.
Plant-stable models (counts)
Controller
Cart-pole (ODE)
Cart-pole (MuJoCo)
Double cart-pole
Double pendulum
Oracle linear MPC
1/1
1/1
1/1
1/1
Neural model + NMPC
10/10
10/10
1/10
0/10
Standard Koopman
0/10
0/10
0/10
0/10
Koopman + invariance penalty
10/10
10/10
0/10
7/10
EDMD ( Korda and Mezić, 2018 )
0/7
0/6
–
9/10
Table 1: Local plant stability at the upright target. Plant-stable models among those with a Schur latent closed loop (Koopman rows) or among evaluated models (oracle MPC, neural NMPC); 10 seeds per learned method, one oracle per system. A dash indicates that no run yielded a stabilizing LQR gain.
Passed estimated margin test
System
Offline data regime
Noise ϵ
Standard Koopman
TIK
Cart-pole
Near target ( 400 pts)
0
0/20
20/20
(upright)
Two scales
0
0/20
20/20
Two scales, noisy states
10−7
0/20
20/20
Closed-loop trajectories
3×10−8
0/20
20/20
Double cart-pole
Two scales
0
0/20
20/20
Table 2: Empirical offline stability-margin tests. Controllers passing the estimated margin test in Assumption 6 ( 20 per method per row). Simulator Jacobians classify zero Standard Koopman and all 20 TIK controllers as locally stable.
Swing-up success (%)
Recovery success (%)
Predictor and terminal set
2 ms
1 ms
0.5 ms
0.25 ms
2 ms
1 ms
0.5 ms
0.25 ms
Fitted linear, estimated Ωc
50
33
25
8
42
47
37
13
Neural, estimated Ωc
91
91
91
–
97
97
97
–
Standard Koopman, latent terminal set
17
15
7
6
18
15
3
4
TIK , estimated Ωc
100
88
88
88
100
99
99
99
Table 3: Dual-mode MPC for constrained cart-pole swing-up and recovery across compute budgets. Share of rollouts ( 12 swing-up and 60 recovery starts per model; 20 TIK and Standard Koopman models, 10 neural, 1 fitted linear) entering their terminal set and stabilizing within force and rail limits. Sets Ωc are estimated using Assumption 15 ; formal guarantees require valid bounds. A dash indicates an exceeded budget.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
System, Koopman model, and controller (Assumption 3 )
x,u,f ; nx,nu
physical state, control input, true dynamics xt+1=f(xt,ut) ; state and input dimensions
(x∗,u∗)
target equilibrium, shifted to the origin (0,0) in the main text
Af,Bf
Jacobians of the true dynamics f at the target equilibrium
Table 4: Summary of notation. Symbols are grouped by the section in which they are introduced. Symbols used only locally within a single proof are defined where they appear.
System
nx
nu
Physical state coordinates
Δt (s)
∣u∣max
Transitions
Simulator
Pendulums
Damped pendulum
2
1
(θ,θ˙)
0.02
8
2.1M
exact (ODE)
Hanging double pendulum
4
2
(θ1,2,θ˙1,2)
0.01
6
2.1M
exact (ODE)
Double pendulum (upright)
4
2
(θ1,2,θ˙1,2)
0.01
20
2.1M
exact (ODE)
Cart-poles
Cart-pole (ODE)
4
1
(x,θ,x˙,θ˙)
0.02
10
2.1M
exact (ODE)
Appendix
Table 5: Summary of benchmark dynamical systems. nx,nu : state and control dimensions; Δt : discrete control period; ∣u∣max : actuator bound per input channel; Transitions: total one-step transitions in the offline training dataset; Simulator: physics engine used for dynamics and ground-truth Jacobians.
Class
Model architecture
Tested / trained
Plant unstable (%) ρ(F(x∗))>1
Linear
Standard Koopman
39/40
100
state
without inverse control loss
40/40
70
recovery
centered at target equilibrium
40/40
100
EDMD, quadratic polynomial dictionary
29/40
100
EDMD, cubic polynomial dictionary
10/40
100
EDMD, thin-plate radial basis functions
15/40
47
Appendix
Table 6: Nominal stability certificate failure rate across extended model classes (four open-loop unstable systems pooled). Tested / trained: models that are nominally certified ( ρ(Acl)<1 ) and admit an unsaturated equilibrium within the training domain, out of all trained models. Unstable on true system (%): percentage of tested models whose deployed loop is physically unstable ( ρ(F(x∗))>1 ).
open-loop
models
certified
deployed, at x∗
simulation
closure
System
ρ(Af)
with x∗
ρ(Acl)
ρ(F(x∗))
unstable
confirms
∥Δ∥/d1
Double cart-pole (MuJoCo)
1.093
22/25
0.9935
1.0672
19 (76%)
21/21
4780
Cart-pole (ODE)
1.083
28/30
0.9863
1.0015
15 (50%)
27/27
18
Cart-pole (MuJoCo)
1.082
30/30
0.9840
0.9979
12 (40%)
30/30
27
Upright double pendulum
1.060
30/30
0.9769
1.0242
27 (90%)
30/30
19
Damped pendulum
0.990
20/20
0.9853
0.9889
5 (25%)
20/20
16
Appendix
Table 7: Breakdown of nominal certificate failures by benchmark system (Standard Koopman and regularized training regimes pooled, excluding the invariance penalty; uncentered controllers). The first four systems are open-loop unstable ( ρ(Af)>1 ). Models with x∗ : models admitting an unsaturated equilibrium out of all trained models (all are nominally certified). Spectral radii and the physical linearization error ratio ∥Δ∥2/d1(F0) (Assumption 9 ) are medians across seeds.
Figure 2: Predicted and deployed spectral radii and stability as the feedback gain is scaled. (a) Spectral radius predicted by the latent model, ρ(Acl)=ρ(A−BK) , plotted against the deployed spectral radius ρ(F) at the nominal LQR gain for Standard Koopman (orange circles), Koopman + invariance penalty (gray squares), and TIK (blue diamonds). (b) Percentage of models ( 40 seeds per method) whose latent stability prediction matches the true deployed stability as the feedback gain is scaled by k∈[0.05,4] .
System
w
n
latent err.
state err.
14-step err.
nonlin.
ρ(Acl)
ρ(F(x∗))
∥E∥2
unstable
Cart-pole (ODE)
0
10
0.526
0.203
0.013
0.006
0.9873
1.0199
134
10/10
Cart-pole (ODE)
0.1
10
0.002
0.061
0.013
0.006
0.9839
0.9842
0.34
0/10
Cart-pole (ODE)
1
10
0.001
0.035
0.012
0.003
0.9839
0.9846
0.21
0/10
Cart-pole (ODE)
10
10
0.000
0.027
0.012
0.001
0.9839
0.9843
0.09
0/10
Cart-pole (MuJoCo)
0
10
0.528
0.217
0.012
0.006
0.9886
1.0210
131
10/10
Cart-pole (MuJoCo)
0.1
10
0.002
0.057
0.012
0.006
0.9840
0.9850
0.33
0/10
Appendix
Table 8: Effect of the soft invariance penalty weight w . Values are medians over n seeds. Setting w=0 corresponds to Standard Koopman. Latent and state errors are one-step prediction errors on held-out validation data; nonlin. measures the relative variation of the encoder Jacobian across the dataset ( 0 for an affine encoder). The final column is the number unstable among models with an admissible equilibrium. The sweep includes runs evaluated at their own equilibria, so these counts differ from the centered comparison in Assumption 1 .
Figure 3: Empirical recovery from initial tilts. Fraction of runs returning to the upright with all poles tilted at rest (two tilt signs, up to 10 seeds per learned curve, one oracle). TIK local feedback uses the fitted physical LQR gain K^ . The lifted TIK controller uses the checkpoint after additional closed-loop training (Assumption F.2 ).
System
Oracle
Before ↑
After ↑
Not worse ↑
Control cells ↓
Cart-pole (ODE)
0.70
0.60
0.70
10/10
0
Cart-pole (MuJoCo)
0.90
0.75
0.75
6/10
6
Double cart-pole
0.15
0.06
0.15
9/9
0
Double pendulum
>1.5
0.40
1.50
10/10
–
Appendix
Table 9: Recovery range of the TIK controller before and after training on closed-loop rollouts. Largest tilt (rad) with at least half of the runs recovered, median over models; “not worse”: models whose range did not shrink; “control”: (model, tilt) cells recovered by the model trained with the same data but no invariance constraint.
System
Offline data regime
Noise ϵ
Bound η contains true error ↑
Curvature bound holds ↑
Cart-pole swing-up
Near the target
0
100/100
80/100
Two scales
0
100/100
100/100
Two scales
10−5
100/100
100/100
Two scales, noisy states
10−7
100/100
100/100
Closed-loop trajectories
3×10−8
100/100
100/100
Double cart-pole
Two scales
0
40/40
40/40
Appendix
Table 10: Offline error bounds checked against simulator derivatives. The study includes the regimes in Assumption 2 and additional noise levels ( 100 fits per cart-pole row and 40 per double cart-pole row).
MPPI only (%)
MPPI + LQR (%)
Model variant
Matched
Bilinear
Entry
Success
Lifted
Fitted
Unstable
Standard Koopman
–
–
95
0
0
91
20/20
Fine-tuned without TIK
–
–
74
0
0
76
20/20
Fine-tuned without TIK
–
✓
100
57
0
99
20/20
TIK, linear input
✓
–
41
6
11
16
0/20
TIK (ours)
✓
✓
100
95
88
99
0/20
Appendix
Table 11: Component ablation on cart-pole swing-up ( 20 models per row). Entry is the fraction of MPPI-only runs reaching the upright neighborhood. Success is the fraction completing the task with MPPI alone. The two MPPI + LQR columns report overall success after switching to lifted LQR or LQR on the fitted physical linearization. The final column counts models with an unstable deployed lifted LQR loop among those with an admissible upright fixed point.
Kinova reach (%) ↑
Six-goal arm (%) ↑
Class
Method
5 ms
20 ms
1 ms
Linear
Linear MPC
100
100
100
Neural
Neural model + MPPI
1
91
0
Neural model + NMPC
–
–
92
Koopman
Standard Koopman
100
100
0
Ours
TIK (ours)
100
100
100
Appendix
Table 12: Real-time MPC on simulated seven-joint manipulators under per-step computation budgets. Success rate (%) on Kinova Gen3 reaching around an obstacle ( 10×30 episodes) and on a torque-controlled arm reaching six goals ( 5×36 episodes). The linear MPC reference uses a fit from logged transitions on Kinova and a single workspace-center linearization on the six-goal arm. The archived protocol gives the reference configuration for this latter baseline and the Go2 baseline, but leaves their Jacobian source unspecified.
Class
Controller
Constraint violations (%) ↓
Stabilized (%) ↑
Linear
Oracle LQR (simulator Jacobians)
2
100
Fitted linear LQR (offline logs)
2
100
Oracle linear MPC (benchmark)
32
100
Neural
Neural model + NMPC
28
91
Neural model + MPPI
7
0
Koopman
Standard Koopman
96
0
Appendix
Table 13: Safe-control-gym cart-pole stabilization benchmark. Percentage of episodes that violate state constraints ( ↓ ) and that successfully stabilize to the target ( ↑ ) across 10 models ×25 initial states.
Success rate (%) at per-step compute budget ↑
Class
Controller
1 ms
2 ms
5 ms
10 ms
20 ms
Linear
Linear MPC (workspace center)
100
100
100
100
100
Neural
Neural model + MPPI
0
1
98
98
98
Neural model + NMPC (relinearized)
92
100
100
97
100
Koopman
Standard Koopman MPC
0
0
0
0
0
Ours
TIK (ours, multi-point QP)
100
100
100
89
97
Appendix
Table 14: Torque-controlled seven-joint manipulator reaching six goals across computation budgets. Success rate (%) over 5 models and 36 episodes per model per budget. TIK’s 10 ms run had deadline misses under machine load. A repeat at lower load reached 100% . The table reports the original budget sweep.
Success rate (%) ↑
Class
Method
Nominal
Noise
Friction
0.8× mass
1.2× mass
Falls (%) ↓
ms/step ↓
Linear
Linear MPC (single standing pose)
3
–
–
–
–
50
0.68
Neural
Neural model + MPPI + filter
47
–
–
–
–
15
6.74
Neural model + NMPC
51
–
–
–
–
4
5.35
Koopman
Standard Koopman QP
26
–
–
–
–
10
0.01
Ours
TIK (ours, multi-point QP)
67
67
73
57
74
21
0.83
Appendix
Table 15: Unitree Go2 quadruped ( 35 D state, 12 D input) posture recovery under pushes and domain shifts ( 150 episodes per entry).