Stochastic Port-Hamiltonian Gaussian Processes for Learning from Noisy Dynamics on Lie Groups
Authors: Thribhuvan Rapolu, Luke McLennan, Chandrajit Bajaj
Organizations: Department of Computer Science, University of Texas at Austin · Oden Institute of Computational Engineering & Sciences, University of Texas at Austin
Learning predictive models for dynamics is difficult because observations are often sparse and noisy, while the underlying dynamics evolve on nonlinear configuration manifolds and are affected by environmental dissipation and stochastic disturbances. Standard neural dynamics models can fit short-term trajectories but often violate physical structure, accumulate error over long horizons, or rely on parameterizations that introduce singularities such as gimbal lock. We introduce LieSPHGP, a stochastic Gaussian-process framework for learning rigid body dynamics with port-Hamiltonian structure and stability-aware robustness. LieSPHGP represents motion on Lie groups and Lie algebras, enabling coordinate-consistent optimization without singular parameterizations. Learned mass, potential energy, dissipation, and control components retain physical roles and a separate diffusion term represents stochastic forcing. Each component is modeled by Matérn kernel Gaussian processes which capture uncertainty induced by noisy observations. To improve physical validity, Lie group exponential map updates preserve rotational constraints, while positive-definite damping ensures nonnegative deterministic energy dissipation. The implicit-explicit integrator additionally provides unconditional stability for the frozen linear damping subsystem. The resulting model learns predictive dynamics that are robust to noise, respect rigid body geometry, and capture both intrinsic motion and environmental dissipation. Experiments on various noisy rigid body benchmarks show improved prediction accuracy, stronger long-horizon stability, and better satisfaction of manifold constraints compared with neural dynamics baselines. Code, data and interactive results are available on the project website: https://thribhuvanrapolu.github.io/liesphgp/
Figures & tables
Figure 1: Overview of the proposed framework. Given noisy observations of a dynamical system, we learn a physics-structured model that captures both the underlying dynamics and their uncertainty while preserving the geometric structure of the state space. The learned model is then used to design a controller that drives the true system toward a desired state or trajectory without additional training. We demonstrate the framework on a 3D pendulum evolving on SO(3) and a quadrotor evolving on SE(3) .
Method
Dynamics Type
Architecture
Geometry
DE Type
NODE ( Chen et al., 2018 )
No Hamiltonian
Neural Network
Euclidean
ODE
HNN ( Greydanus et al., 2019 )
Hamiltonian
Neural Network
Euclidean
ODE
NSDE ( Liu et al., 2019 )
No Hamiltonian
Neural Network
Euclidean
SDE
VIN ( Saemundsson et al., 2020 )
Lagrangian
Neural Network
Euclidean
ODE
PHNN ( Desai et al., 2021 )
Port-Hamiltonian
Neural Network
Euclidean
ODE
InfGAN ( Kidger et al., 2021 )
No Hamiltonian
Neural Network
Euclidean
SDE
Table 1: Summary of references by dynamics type, architecture, geometry, and differential equation type, sorted by year.
Figure 2: LieSPHGP framework. Five GP subnetworks ( M−1 , V , D , G and the diffusion ξ ), each a level (constant, or linear for V ) plus a Matérn-kernel random-feature Gaussian process (Section 3.2 , Algorithm 5 ), define a stochastic port-Hamiltonian system on SO(3) or SE(3) (Section 3.1 ), integrated with the Lie-IMEX solver (Section 3.3 , Algorithm 4 ). At each training step one GP weight sample W∼q(W) is drawn and held fixed throughout each rollout; starting from y0 , the solver gives the prediction x^k− and its Jacobians Ak , Bk , and a Lie-group EKF propagates the covariance Pk−=AkPk−1Ak⊤+BkBk⊤ , so the diffusion enters only through the process covariance BkBk⊤ , while the noisy observations yk and the learned measurement-noise scales σobs enter the measurement update. Training minimizes L=Eq(W)[EKF NLL]+βKL(q(W)∥p(W)) , the EKF negative log-likelihood of the noisy windows plus a KL term on the GP weights (Section 3.4 , Eq. 21 ). Each parameter group has its own Prodigy optimizer, so its learning rate ηk(t) is adapted separately during training (Algorithm 1 ).
Figure 3: Open-loop prediction on the windy pendulum. Roll, pitch, and yaw (Euler xyz angles of R , as in Figure 4 ) predicted over 10 s from the true initial state using recorded inputs. Roll =±π corresponds to the hanging position. Black shows ground truth. Lines show the mean of 32 sampled paths, except for PH-NODE’s single deterministic path. Shading shows ±1σ and ±2σ . The true SDE uses independently sampled wind. We show the held-out trajectory with the largest difference in mean rotation error between Lie-PH-NN-SDE and LieSPHGP ( 43.1∘ versus 7.0∘ ). Over all 10 held-out trajectories, Lie-PH-NN-SDE has the larger error on 9 . On the shown trajectory, 2σ coverage is 0.97 for both the true SDE and LieSPHGP, and 0.47 for Lie-PH-NN-SDE.
5-point (0.2 seconds) windows
41-point (2 seconds) windows
Error
PH-NODE
LieSPHGP ( ξ=0 ) (Ours)
PH-NODE
LieSPHGP ( ξ=0 ) (Ours)
Rotation (rad)
1.381±0.016
1.064±0.067
0.773±0.083
0.066±0.017
Velocity (rad/s)
0.872±0.004
0.611±0.017
0.636±0.034
0.048±0.007
Energy
1.154±0.007
0.787±0.037
0.714±0.033
0.044±0.005
Det. ( ×10−6 )
66.0±0.1
0.98±0.03
51.7±2.8
0.99±0.03
Orth. ( ×10−6 )
84.3±0.2
1.92±0.05
72.3±4.4
1.93±0.07
Table 2: Open-loop prediction and identification on the noisy deterministic SO(3) pendulum. Constant friction 0.5 , observation noise 0.25 , no diffusion ( ξ=0 for LieSPHGP). Both methods are trained on the same data in two styles: PH-NODE’s original 5-point windows (stride 1, full batch) and 41-point windows (stride 30, batch 64). PH-NODE is the published model and loss ( Duong et al., 2024 ) with M−1→I pretraining. Errors over 10 -s rollouts of 10 held-out trajectories, mean ± std over 5 training seeds; best in bold. With the same windows, LieSPHGP has the lower mean error on every row.
Error
PH-NODE
NeuralSDE
Lie-PH-NN-SDE
LieSPHGP
Rotation (rad)
1.066±0.038
1.78±0.10
0.471±0.047
0.291±0.025
Velocity (rad/s)
0.776±0.007
50.2±3.7
0.538±0.015
0.443±0.012
Energy
0.703±0.010
(3.1±0.5)×103
0.385±0.014
0.320±0.011
Det.
(5.2±0.2)×10−4
(4.9±1.2)×108
(8.8±0.1)×10−7
(9.9±0.1)×10−7
Orth.
(6.1±0.2)×10−4
(2.9±0.3)×106
(1.71±0.02)×10−6
(1.91±0.01)×10−6
Rotation CRPS
0.443±0.014
0.887±0.049
0.174±0.019
0.100±0.010
Table 3: Open-loop prediction and identification on the stochastic windy SO(3) pendulum. Rate-dependent friction, wind ξ=0.5 , observation noise 0.5 . Errors over 10 -s rollouts of 10 held-out trajectories, mean ± std over 5 training seeds. Stochastic models are scored on 32 sampled paths with an independent wind draw; CRPS and 2σ coverage measure the predictive distribution. The lower block is the relative error of the learned port-Hamiltonian products. PH-NODE is the published model in Duong et al. (2024) . NeuralSDE is an unstructured neural SDE ( Liu et al., 2019 ) in Euclidean coordinates. LieSPHGP and Lie-PH-NN-SDE differ only in function parameterization (GP versus neural networks) and in the weights (variational posterior versus point estimate). They share the optimizer, the state-dependent diffusion and all other modeling and numerical settings. Rotation error is the geodesic angle arccos((tr(R⊤R^)−1)/2) . For sampled models, R^ is the chordal mean, obtained by projecting the arithmetic mean of the sampled rotation matrices onto SO(3) using SVD. For single-path models, R^ is the predicted rotation. Rotation CRPS (rad) and 2σ coverage use the tangent coordinates Log(R^⊤R) at R^ and angular-velocity CRPS is in rad/s. The SO(3) constraint errors are computed on the raw, unprojected rollouts.
Figure 4: Closed-loop IDA-PBC control of the stochastic SO(3) pendulum using learned dynamics. Each model supplies its learned dynamical subnetworks to the same IDA-PBC controller acting in the windy pendulum environment, with the objective of stabilizing the pendulum at the upright configuration. LieSPHGP (orange) closely follows the controller constructed from the true model, which uses ground truth components (green), and reaches the target, whereas PH-NODE (purple) does not reach the upright. The experiment demonstrates that differences in learned physical structure that are modest in open-loop prediction can become consequential for downstream control.
Gain scale
PH-NODE
LieSPHGP (Ours)
True model
Success (%)
s=1
31±9
100±0
100
s=0.25
0±0
78±5
84
Final angle (rad)
s=1
1.67±0.33
0.027±0.001
0.024
s=0.25
2.06±0.16
0.088±0.004
0.071
Table 4: IDA-PBC swing-up of the windy pendulum at normal and low controller gains. Gains KR=25s , Kω=10s . Success: angle to upright stays below 10∘ ( 0.175 rad) over the last 2 s of the 10 -s run. Final angle: angle to upright at t=10 s. Mean ± std over 5 training seeds, each with 5 initial attitudes ×5 Brownian motion samples. True model uses the ground-truth dynamics in the same controller.
Figure 5: Control using the learned dynamics. The learned subnetworks M−1 , V , D , and G define the IDA-PBC control law. Energy shaping replaces the learned potential with a desired potential whose minimum is at the target, while damping injection adds damping Kω . The errors are Lie-group differences, xe=x⊖x∗ (Section 3.4 ), whose attitude part is eR=Log(R∗⊤R)∨ with R∗ the target attitude. For the quadrotor, IDA-PBC forms the inner attitude loop of an outer position-tracking controller (Appendix A.4 ).
Metric
PH-NODE
LieSPHGP (Ours)
True model
Position RMSE (m)
0.251±0.005
0.133±0.009
0.127
Linear velocity RMSE (m/s)
0.856±0.015
0.485±0.018
0.464
Attitude RMSE (rad)
0.220±0.006
0.203±0.003
0.201
Angular velocity RMSE (rad/s)
4.18±0.06
3.93±0.04
3.89
Energy RMSE
0.027±0.001
0.016±0.000
0.016
Valid tracking time (s)
8.39±0.17
9.22±0.14
–
Table 5: Closed-loop tracking using learned SE(3) quadrotor dynamics. Each learned model supplies the drift and input map of the same geometric tracking controller with an IDA-PBC attitude loop (Appendix A.4 ), which commands the four rotor thrusts, flying 10 held-out 10 -s references under 5 Brownian motion samples on the PyBullet plant. Values are mean ± std over 3 training seeds. True model uses the ground-truth components in the same controller. LieSPHGP nearly matches the true-model controller and halves PH-NODE’s tracking error. Position, linear-velocity and angular-velocity RMSE are measured against the reference and the attitude error is the RMS geodesic angle between the flown attitude and the attitude Rd commanded by each model’s controller. Valid tracking time is the first time the position leaves the true-model flight under the same wind sample by more than 0.158 times the RMS distance of the reference from its starting point. Tracking CRPS compares the states reached under five wind realizations with the deterministic reference. It measures tracking performance under stochastic disturbances and does not assess predictive calibration.
Figure 6: Closed-loop SE(3) quadrotor tracking using learned dynamics. Three-dimensional trajectories for two of the ten held-out 10 -s flights (A and B). Thin lines show individual Brownian motion samples, thick lines show the mean over those five samples, and dashed lines show the reference. Each model supplies the dynamics to the same geometric tracking controller with an IDA-PBC attitude loop. The controller’s commands are allocated to the four rotors of the contact-free PyBullet CF2P plant ( Panerati et al., 2021 ) . LieSPHGP (orange) tracks close to the controller using the true dynamics (green), while PH-NODE (purple) deviates during turns. Across all flights and seeds, the position RMSE is 0.133 m for LieSPHGP and 0.251 m for PH-NODE (Table 5 ).
Metric
PH-NODE
LieSPHGP (Ours)
True model
Position error (m)
0.431±0.009
0.121±0.011
0.103
Attitude error (rad)
0.313±0.004
0.174±0.008
0.159
Angular velocity error (rad/s)
1.995±0.026
1.165±0.039
1.040
Position CRPS (m)
0.199±0.006
0.043±0.003
0.037
Position 2σ coverage
0.00
0.66±0.02
0.69
Thrust difference (/ hover)
0.360±0.019
0.192±0.004
0.175
Table 6: Model-in-the-loop prediction on the quadrotor. The controller of Table 5 , built from the true dynamics, computes the rotor thrusts from the learned model’s own predicted state, and the learned model (with its learned diffusion, 8 paths) is propagated as the plant; the prediction is compared with the same controller flying the real PyBullet plant ( 10 held-out references ×5 Brownian motion samples). Errors averaged over 0 – 10 s; mean ± std over 3 training seeds. True model: the ground-truth SDE as the plant with an independent Brownian motion sample.
Figure 7: Open-loop prediction on a held-out BlueROV2 test segment. Predicted positions x , y , and z (positive downward), and position error, using recorded thruster commands and the measured initial state. Vertical dotted lines mark the first time position error exceeds 0.1 m (valid prediction time). LieSPHGP shows the mean of 32 sampled paths with a ±2σ band (training seed 0 ). The physics model uses published vehicle parameters. LieSPHGP learns M−1 , V , D , and G from training data alone. The shown segment has the median LieSPHGP valid prediction time among the 15 test segments.
Model
T=1 (0.02 s)
T=10 (0.2 s)
T=100 (2 s)
VPT (s)
Persistence
0.091
0.122
0.402
0.94
Physics (published parameters)
0.086
0.128
0.301
1.62
Koopman (EDMDc)
0.096
0.126
0.259
0.81
LieSPHGP (ours)
0.090±0.000
0.105±0.002
0.252±0.034
1.98±0.28
Table 7: Prediction on real BlueROV2 trajectories. Endpoint RMSE across the 12 state variables (position, Euler angles, body linear and angular velocities) after T steps of 0.02 s. As in the benchmark, the RMSE pools all 12 components with their native units (m, rad, m/s, rad/s). Predictions start from each of the 9,164 test samples in the Marinarium recording ( Torroba et al., 2026 ) and use the recorded thruster commands. Lower RMSE is better. Valid prediction time (VPT) is the time until the position error first exceeds 0.1 m, averaged over 15 held-out 10 -s segments. Higher VPT is better. Koopman (EDMDc) uses the code and hyperparameters of Torroba et al. (2026) and the same training trajectories as LieSPHGP. Physics uses the published BlueROV2 parameters. LieSPHGP uses the rollout with posterior-mean GP weights (no process noise) for endpoint RMSE and the mean of 32 sampled paths for VPT. Its results are reported as mean ± standard deviation over two training seeds.
System
Metric
Embedded
Intrinsic
Pendulum
Test EKF NLL
0.7875
0.7875
Rotation error (rad)
0.259±0.13
0.266±0.13
Rotation CRPS
0.091±0.038
0.093±0.040
Damping error
0.861
0.880
Training time (min)
15.8
25.6
Quadrotor
Test EKF NLL
0.1137
0.1137
Table 8: Embedded vs. intrinsic SO(3) kernel. Same data, F=64 features, training seed 0, learned kernel hyperparameters and 5,000 training steps; only the features of the GPs with rotational input change. Pendulum errors are 0 – 10 s open-loop values and the quadrotor position RMSE is the closed-loop tracking error, as in the main tables (mean ± std over held-out trajectories). Damping error: relative error of M−1D along the ground-truth closed-loop states (protocol of Tables 12 and 16 ).
Determinant
Orthogonality
Model
Lie-IMEX
RK4
Lie-IMEX
RK4
Deterministic pendulum
LieSPHGP ( ξ=0 )
(9.8±0.5)×10−7
(3.3±1.7)×10−5
(1.9±0.08)×10−6
(5.1±2.8)×10−5
PH-NODE
(9.6±3.8)×10−7
(6.1±1.3)×10−5
(2.1±0.3)×10−6
(7.8±1.1)×10−5
Windy pendulum
LieSPHGP
(9.3±0.8)×10−7
(8.9±0.4)×10−5
(1.8±0.1)×10−6
(1.0±0.04)×10−4
Table 9: SO(3) constraint error of the same trained models under two integrators. Each trained pendulum model is rolled out over 10 s with Lie-IMEX and with RK4 at the same step h=0.05 s (no retraining). Values: mean ± std over the 10 held-out trajectories of maxt∣detRt−1∣ and maxt∥Rt⊤Rt−I∥F , computed in float64 from float32 rollouts. RK4 rollouts carry no diffusion.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: 3D Windy Rigid-Body Pendulum
Figure 9: SE(3) windy quadrotor. The four rotor thrusts F1,…,F4 are the inputs; gravity, velocity-dependent air drag −Dv(v)v and random wind forces and torques ξdWt act on the body.
Model
p Error (m)
v Error (m/s)
R Error (rad)
ω Error (rad/s)
Naïve ( Busetto et al., 2026 )
0.3536
0.7745
0.1639
0.5797
Physics ( Busetto et al., 2026 )
0.0448
0.2094
0.1014
0.5797
Res-MLP ( Busetto et al., 2026 )
0.0795
0.2488
0.1210
0.7097
Hybrid ( Busetto et al., 2026 )
0.0456
0.2051
0.1228
0.5782
Res-LSTM ( Busetto et al., 2026 )
0.1182
0.3314
0.1129
0.7149
PH-NODE ( Duong et al., 2024 )
0.0393
0.2262
0.1451
0.7849
Appendix
Table 10: Comparison on the real-world quadrotor dataset from Busetto et al. (2026) . Multi-step prediction MAE averaged over rollout horizons 1 through 50 (step size 0.01 s, i.e. up to 0.5 s) from every start state of the held-out Melon flights ( 19,347 windows), scored per flight as in Eq. (11) of Busetto et al. (2026) . Baselines: the benchmark’s released predictions rescored by the same script. The published values pool the three flights (e.g. Physics 0.0464 , 0.2125 , 0.1020 , 0.5817 ) and give the same ranking. LieSPHGP: predictions with posterior-mean GP weights of a single training run. PH-NODE ( Duong et al., 2024 ) : a single training run. Bold: best in each column.
Figure 10: Real BlueROV2 Heavy underwater vehicle with eight thruster inputs u1,…,u8 . Gravity acts at the center of gravity (CG), and buoyancy acts at the center of buoyancy (CB). Water contributes hydrodynamic drag −D(ν)ν , added mass MA , and the associated Coriolis term CA(ν) . The body frame follows the marine convention, with zb pointing down.
Model
T=10 ( 0.2 s)
T=100 ( 2 s)
Position 0 – 3 s (m)
VPT (s)
Persistence
0.218
0.397
0.479
0.51
Physics (published parameters)
0.359
0.673
0.282
1.19
Koopman (EDMDc)
0.331
0.844
0.544
0.58
LieSPHGP (ours)
0.226±0.003
0.412±0.007
0.197±0.009
1.13±0.03
Appendix
Table 11: BlueROV2: evaluation on unseen trajectories. Models trained on the manual recording are evaluated, without retraining, on the attitude-stabilized recording of Torroba et al. (2026) ( 45,852 test samples, 60 segments of 10 s). Endpoint RMSE as in Table 7 ; position error averaged over 0 – 3 s; VPT as in Table 7 . LieSPHGP: mean ± std over two training seeds.
Figure 11: Physics Identification in Windy Pendulum System: Comparison of the learned diagonals of the matrices M−1G for control (top row) and M−1D for damping (bottom row) along the ground-truth swing-up flight (training seed 0 ). The models compared are the ground truth (black, dashed), PH-NODE ( Duong et al., 2024 ) (purple), Lie-PH-NN-SDE (blue), and LieSPHGP (orange). LieSPHGP follows the ground truth for both products; PH-NODE over-estimates the damping about six times, and Lie-PH-NN-SDE has the wrong sign of the y control gain near the upright.
Product
LieSPHGP (ours)
Lie-PH-NN-SDE (ours)
PH-NODE (prior work)
Damping M−1D
0.44±0.11
0.67±0.11
4.96±0.07
Control gain M−1G
0.057±0.003
0.392±0.024
0.347±0.009
Gravity aV
0.054±0.011
0.285±0.040
0.402±0.006
Diffusion ξ
0.20±0.07
0.30±0.05
–
Appendix
Table 12: Relative error of the learned pendulum physics: mean∥Q−Q∥/mean∥Q∥ over the states of the 10 held-out trajectories ( Q : ground truth, Q^ : learned), mean ± std over 5 training seeds, bold = best learned model. Diffusion: relative error of the learned diagonal of ξ against (0.5,0.5,0) .
Quantity
Truth
LieSPHGP (ours)
Deterministic SO(3) pendulum
Obs. noise (σR,σω)
(0.25,0.25)
(0.247,0.252)
NIS per dimension (ideal 1 )
1
1.00±0.01
Windy SO(3) pendulum
Obs. noise (σR,σω)
(0.5,0.5)
(0.495,0.500)
Diffusion ξ
(0.5,0.5,0)
(0.61,0.57,0.06)
Appendix
Table 13: Learned noise against the simulators’ ground truth. The EKF likelihood learns the observation noise σobs and the process noise ξ as separate quantities. Mean over 5 training seeds. NIS is normalized innovation squared of the EKF on the noisy test windows, per dimension (ideal 1 ).
Figure 12: Calibration of the predictive distribution on the windy pendulum. 2σ coverage of rotation (left) and angular velocity (middle), i.e. the share of true states inside the predicted ±2σ band, and the spread-skill ratio MSE/variance of rotation (right, log scale, 1 = calibrated, >1 = over-confident), averaged over 0 – 1 , 0 – 3 , 0 – 5 and 0 – 10 s of open-loop prediction on the 10 held-out trajectories. 32 sampled paths per model and the true model is the ground-truth SDE with an independent wind draw. Mean ± std over training seeds. Dotted lines: nominal 0.954 and 1 .
Variant
Rotation
Rot. CRPS
2σ cov.
M−1G
M−1D
Swing-up
LieSPHGP
0.288±0.024
0.099
0.91
0.057
0.43
100 / 81
Loss
Rollout MSE, noisy start
0.628±0.175
0.248
–
0.807
2.96
0 / 0
Rollout MSE, learned start
0.308±0.024
0.122
0.27 – 0.43
0.187
0.32
100 / 66
GP weights
Point estimate, MLE
0.512±0.030
0.191
0.60
0.359
0.94
9 / 0
Appendix
Table 14: Ablations on the windy pendulum. Each variant changes one ingredient of LieSPHGP, except that the rollout-MSE variants, whose loss has no noise model, are also trained without the diffusion and the observation-noise model on same data, 5,000 steps, final checkpoint. Mean ± std over 3 training seeds and swing-up success at gain scales s=1 / 0.25 .
Metric
SPH-NN
SPH-NN-mb
LieSPHGP (ours)
Rotation error (rad)
1.71±0.03
1.75±0.09
0.288±0.024
Velocity error (rad/s)
4.95±1.21
5.34±0.99
0.441±0.013
Energy error
33±20
27±15
0.317±0.013
maxt∥Rt⊤Rt−I∥F
347±105
73±30
1.9×10−6
Rotation CRPS
0.61±0.05
0.69±0.08
0.099±0.010
Rotation 2σ coverage
0.80±0.11
0.56±0.05
0.91±0.03
Appendix
Table 15: SPH-NN and LieSPHGP on the windy pendulum. The data and evaluation protocol follow Table 3 . Values with error bars are means and standard deviations over the same three training seeds ( 0 – 2 ). Swing-up success is reported at normal and low IDA-PBC gains (Table 4 ).
Figure 13: Physics Identification in Quadrotor System (Control Matrix): Actuated rows of the learned control map M−1G along a 10 -s ground-truth tracking flight (training seed 0 ). The top row shows body- z thrust ( vz ), and the next three show body torques ( ωx,ωy,ωz ), with one column per rotor. The torque rows contain the rotor mixing matrix Bmix . Each row is divided by the magnitude of its nonzero ground-truth entries ( 37.0 , 1658 , 1658 , and 777 , from top to bottom). The normalized ground truth (dashed black) is therefore −1 , 0 , or 1 , and all panels share the same vertical scale. PH-NODE ( Duong et al., 2024 ) is shown in purple and LieSPHGP in orange. The body- x and body- y force rows are omitted because they are zero in the true model and the controller does not use them.
Figure 14: Physics Identification in Quadrotor System (Dissipation): Comparison of the learned diagonals of M1−1Dv for translational damping (top row) and M2−1Dω for rotational damping (bottom row) along a 10 -s ground-truth tracking flight (training seed 0 ). The models compared are the ground truth (black, dashed), PH-NODE ( Duong et al., 2024 ) (purple), and LieSPHGP (orange). LieSPHGP is closer to the ground truth for both translational and rotational damping, while PH-NODE over-estimates both (Table 16 ).
Product
LieSPHGP (ours)
PH-NODE (prior work)
Thrust row of M1−1G
0.080±0.004
0.182±0.017
Torque rows of M2−1G ( Bmix )
0.017±0.003
0.014±0.001
Gravity −M1−1RT∇V
0.066±0.010
0.355±0.024
Translational damping M1−1Dv
0.38±0.19
7.99±0.42
Rotational damping M2−1Dω
0.46±0.22
2.61±0.03
Diffusion ξ
0.050±0.039
–
Appendix
Table 16: Relative error of the learned quadrotor physics: mean∥Q−Q∥/mean∥Q∥ over the states of the held-out flights ( Q : ground truth, Q^ : learned), mean ± std over 3 training seeds; bold = best learned model. The control map M−1G ( 6×4 , one column per rotor) is split into its thrust row (body z ) and its 3×4 torque rows, which contain the rotor mixing matrix Bmix . Diffusion: relative error of the six learned entries of ξ against 0.5 .
5-sample windows
101-sample windows
Metric
PH-NODE
LieSPHGP
PH-NODE
LieSPHGP
True model
Closed-loop position RMSE (m)
0.255
0.199
0.745
0.130
0.127
Closed-loop velocity RMSE (m/s)
0.867
0.692
1.615
0.479
0.464
Rotor saturation share
0.103
0.072
0.296
0.048
0.040
Model-in-the-loop position error (m)
0.422
0.163
0.453
0.113
0.103
Model-in-the-loop 2σ coverage
0.00
0.47
0.00
0.65
0.69
Appendix
Table 17: Effect of the training data style on the quadrotor. Both methods trained on the same rotor-input data with PH-NODE’s style ( 5 -sample windows, stride 1 , batch 256 ) and with LieSPHGP’s style ( 101 -sample, i.e. 2 -s, windows, stride 75 , batch 16 ); everything else is each method’s own recipe ( 5,000 steps, final checkpoint). Training seed 0 ; closed loop and model-in-the-loop over 10 held-out references ×5 Brownian motion samples, 0 – 10 s. Lower block: relative identification errors.
x^t−,(At,Bt)←
LieIMEXSolver(x^t−1,ut(j),0;θ(W))
Appendix
Algorithm 2 Training Lie Group Stochastic Port-Hamiltonian Gaussian Process (Detailed)
Dynamical systems studies how states evolve over time, including stability, feedback, chaos, and transitions in physical, biological, and engineered systems.
In this paper, we propose Lagrangian Gaussian Processes (LGPs) for probabilistic and data-efficient learning of dynamics via discrete forced Euler-Lagrange equations. Importantly, the geometric structure of the Lagrange-d'Alembert principle, which governs the motion of dynamical systems, is preserved by construction in the absence of external forces. This allows learning physically consistent models that overcome erroneous drift in the system's energy, thereby providing stable long-term predictions. At the core of our approach lie linear operators for Gaussian process conditioning, constructed from discrete forced Euler-Lagrange equations and variational discretization schemes. Thereby and unlike prior work, the method enables learning dynamics from discrete position snapshots, i.e., without access to a system's velocities or momenta. This is particularly relevant for a large class of practical scenarios where only position measurements are available, for instance, in motion capture or visual servoing applications. We demonstrate the data-efficiency and generalization capabilities of the LGPs in various synthetic and real-world case studies, including a real-world soft robot with hysteresis. The experimental results underscore that the LGPs learn physically consistent dynamics with uncertainty quantification solely from sparse positional data and enable stable long-term predictions.
We propose an architecture for learning the dynamics of mechanical systems based on discrete forced Euler-Lagrange equations on Lie groups using only position data. By formulating the dynamics directly on manifold-valued configuration spaces, the method naturally respects the geometric structure of the systems and preserves geometric invariants and conservation laws. The reliance on position measurements alone makes the framework applicable in settings where velocity data are unavailable or noisy. The approach extends naturally to multibody systems, accommodates external control inputs, and demonstrates strong performance on both synthetic and real-world datasets.
Martine Dyring Hansen, Marta Ghirardelli, Elena Celledoni +2
Department of Mathematics and Cybernetics, SINTEF Digital Forskningsveien 1, 0373 Oslo, Norway · Department of Mathematical Sciences Norwegian University of Science and Technology (NTNU) Alfred Getz’ vei 1, 7034 Trondheim, Norway · Instituto de Ciencias Matematicas (CSIC) Calle Nicolás Cabrera 13-15, Cantoblanco, 28048 Madrid, Spain
Structure-preserving neural networks are essential for the long-term prediction of Hamiltonian systems from data. Many important Hamiltonian systems in mechanics and control admit symmetry reduction to Lie--Poisson systems, including rigid bodies, underwater vehicles, fluids, plasmas, and optimal control problems. A fundamental challenge in learning such systems is that their dynamics evolve in momentum variables that are typically unobservable, while available data consist only of observable quantities such as configurations and velocities. In optimal control applications, the situation is further complicated because the latent variables contain unobservable co-states and the Hamiltonian may be degenerate, preventing the existence of a corresponding Lagrangian and rendering the encoder-decoder approaches inapplicable. We introduce Latent Lie--Poisson Neural Networks (LLPNNs), a structure-preserving framework for learning Lie--Poisson dynamics directly from observable data. The proposed approach exploits three geometric ingredients: (i) learning either a Hamiltonian decoder or a pseudo-Lagrangian encoder on the active variables, (ii) constructing latent trajectories through a universal Noether invariant arising from Lie--Poisson symmetry reduction, and (iii) reconstructing observable and latent dynamics through Lie--Poisson flows combined with Magnus-based Lie-group updates. The resulting method preserves the geometric structure and is applicable to both regular and degenerate Hamiltonian systems. We demonstrate the method on three examples: a generalized rigid body on SO(3), Kirchhoff's underwater vehicle on SE(3), and an optimal-control problem for interacting vehicles on SE(2)N. Numerical experiments show excellent long-term predictive accuracy, strong robustness to noise, and competitive performance using only modest datasets and lightweight neural-network architectures.
Vakhtang Putkaradze
Department of Mathematics, The University of Alabama, Gordon Palmer Hall, Tuscaloosa, AL, 35401, USA