At a glance
Neural Contractive Dynamical Systems (NCDS) learn motion fields by parameterizing a negative-definite matrix field and integrating it to obtain velocities. An injective variational autoencoder supports low-dimensional latent dynamics, and a restricted rotation representation extends the method to full poses. Experiments include handwriting, robot trajectories, and human motion. The paper’s stability claims concern its dynamical-system construction; decoded contraction applies on the learned manifold, with separate treatment of off-manifold initialization.
1 Stability in robot learning
Convergence to a goal does not by itself ensure that nearby motions approach one another. The authors use contraction to constrain extrapolation away from demonstrations, contrasting their architecture with mixture models, kernels, and separately learned contraction metrics. The aim is to fit flexible motion models with ordinary optimization while controlling their stability properties. [2], [31], [42].
2 Learning contractive vector fields
The method begins with a differential condition for contraction and then builds a neural motion model around it.
2.1 Background: contractive dynamical systems
Uniform negativity of the symmetric part of a vector field’s Jacobian makes nearby trajectories converge exponentially in the Euclidean metric. More general contraction metrics permit state-dependent geometry; the base NCDS construction targets an identity metric.
Equations 1–2
| (1) |
| (2) |
2.2 Neural contractive dynamical systems (NCDS)
A network produces a matrix whose negative Gram matrix, shifted by a positive diagonal margin, is negative definite. The authors integrate this matrix field along a straight segment from a learned reference state to predict velocity, presenting the result as a contractive neural architecture. Training fits demonstrated velocities or next states through numerical integration.
Equation 3
| (3) |
Equations 4–5
| (4) | ||||
| (5) |
3 Learning latent contractive dynamics
High-dimensional models can be expensive and brittle with limited demonstrations. The latent variant learns a smaller motion space while constraining its decoder geometry.
3.1 Background: deep generative models
Variational autoencoders provide latent representations; injective flows supply the required one-to-one decoder structure.
Variational autoencoders (VAEs).
The VAE encodes observed states into a lower-dimensional distribution and learns a decoder by maximizing an evidence lower bound. Latent states serve as the inputs to the motion model.
Injective flows.
Zero-padding followed by invertible transformations embeds a low-dimensional latent space into a higher-dimensional observation space without self-intersections.
3.2 Latent neural contractive dynamical systems
The injective decoder is a diffeomorphism onto its image, allowing the paper’s contraction argument to transfer to that learned manifold. Its Jacobian maps latent velocities into observation-space velocities. Off-manifold initial states are projected through an approximate encoder; the initial movement to the manifold is explicitly not guaranteed to be contractive.
Equations 6–8
| (6) |
| (7) |
| (8) |
3.3 Learning position and orientation dynamics
Full-pose motion combines Euclidean positions with rotation geometry.
Orientation parameterization.
Rotation matrices lie on SO(3). Logarithmic and exponential maps relate them to three-dimensional rotation vectors, but wrapping prevents a globally one-to-one parameterization. The method restricts vectors to norms below π.
NCDS on Lie groups.
An injective decoder constrained to the open π-ball is followed by the rotation exponential map. Position and rotation coordinates can then share latent dynamics while retaining the stated coordinate-change argument within this restricted representation.
3.4 Obstacle avoidance via matrix modulation
Obstacle geometry and position define a modulation matrix that reshapes the decoded velocity field near obstacles. The paper invokes prior modulation theory for stability preservation. The demonstrated construction avoids obstacles at the end effector; whole-body avoidance requires additional modeling. [12].
4 Experimental results
Evaluation emphasizes reconstruction accuracy, nearby-trajectory behavior, higher-dimensional representations, and qualitative robot recovery.
Datasets.
Ten LASA handwriting shapes are selected and preprocessed to share endpoint conventions. Stacking shapes creates four- and eight-dimensional datasets. Further data include five seven-joint robot trajectories and kinesthetic end-effector demonstrations.
Metrics.
Reconstruction is measured using the paper’s stated DTWD formula, which sums nearest-point distances in both trajectory directions. Nearby-trajectory distance over time is also plotted to assess contraction behavior.
Equation
Baseline methods.
Euclideanizing Flow, Imitation Flow, and SEDS are comparison methods with asymptotic stability objectives; they do not target the same contraction guarantee.
4.1 Comparative results
NCDS is not most accurate in low dimensions: Euclideanizing Flow wins on LASA-2D and Imitation Flow on LASA-4D. NCDS has the lowest reported DTWD on LASA-8D (2.28 ± 0.24) and seven-joint data (1.18 ± 0.16). The latent representation is central to these higher-dimensional gains.
4.2 Obstacle avoidance
Handwriting fields and a robot demonstration show trajectories rerouting around a known obstacle while reaching the target. These are controlled demonstrations of the modulation method, not a broad collision-safety benchmark.
4.3 Robot experiment
A Franka arm reproduces taught sinusoidal and V-shaped motions using Cartesian impedance control, including manually perturbed executions. Additional experiments cover synthetic position-orientation data, simulated joint-space reproduction, and human-motion trajectories.
5 Discussion
The architecture combines a constrained matrix-field model, latent representation, orientation handling, and optional obstacle modulation. The evidence supports useful motion reproduction in the tested settings, with numerical and representation assumptions limiting the scope of its guarantees.
Limitations.
Results are sensitive to the integration scheme. Adaptive stepping improves reliability but adds computation and variable latency; high-dimensional queries may not meet strict real-time requirements.
Acknowledgments
The work acknowledges VILLUM FONDEN, the European Research Council, and the Novo Nordisk Foundation.
Appendix A Appendix
The appendix retains mathematical details, algorithms, architectural choices, and additional comparisons.
A.1 Negative Definiteness of a Symmetric Matrix
The appendix relates a symmetric matrix’s negative-definite quadratic form to its eigenvalues, supporting the matrix parameterization used in the method.
A.2 Lie operators of 𝒮 𝒪 ( 3 )
Closed-form logarithmic and exponential maps specify the rotation representation and its angle-dependent terms.
Equations 9–10
| (9) | ||||
| (10) |
A.3 Obstacle avoidance via matrix modulation
The modulation matrix changes radial and tangential velocities according to a differentiable obstacle-distance function and reactivity parameter. Known obstacle geometry is required, and the authors cite the prior method’s stability analysis.
Equations 11–12
| (11) |
| (12) |
A.4 Implementation Details
The injective decoder uses three rational-quadratic spline coupling layers with small residual networks. The Jacobian network has two 500-unit hidden layers. PyTorch and torchdiffeq implement training and numerical integration.
Loss functions:
The VAE uses the evidence lower bound, including a Kullback–Leibler term. The motion model then minimizes latent next-state reconstruction error.
Equations 13–14
| (13) |
| (14) |
Algorithms:
Training occurs in stages: fit the VAE, encode demonstrations, estimate latent velocities, and fit the NCDS model. At execution, encode the pose, predict latent velocity, and apply the decoder Jacobian. Both original algorithms are retained.
A.5 An injective decoder architecture over the π -ball ℬ π
A tanh layer and a box-to-ball transformation constrain decoder outputs to the π-ball before applying the rotation exponential map.
Equations 15–16
| (15) |
| (16) |
A.6 Additional Experiments
Additional experiments examine richer poses, human motion, numerical choices, and baseline scaling.
A.6.1 Learning ℝ 3 × 𝒮 𝒪 ( 3 )
Synthetic full-pose trajectories combine different LASA shapes for position and orientation. Quaternion-derived rotations are mapped to the Lie algebra, producing a six-coordinate state whose learned latent trajectories are visualized.
A.6.2 Learning human motion
A tennis forehand from the KIT motion dataset is represented by 44 joint coordinates and compressed to two latent dimensions. Generated human-dummy motions illustrate reproduction and extrapolation from different starts.
A.6.3 Quantitative results
Larger contraction margins reduce expressiveness in the reported ablation. Unconstrained matrix, MLP, and neural-ODE baselines fit observed regions but lack the same observed stability. Integration tolerances and activations affect behavior. Reported single-step timings are 40.9 ms for direct eight-dimensional modeling, 20.2 ms with latent modeling, and 121.0 ms for 44-dimensional latent modeling.
A.6.4 Comparisons
Baseline implementations are tuned for reconstruction: flow architectures vary in depth and coupling structure, while SEDS uses Gaussian components. The comparison assesses accuracy and general stability because the methods offer different stability properties.
A.6.5 Synthetic LASA
The synthetic benchmark grows from two to eight dimensions by concatenating distinct handwriting motions. Baseline capacity is increased and selected from multiple configurations; NCDS keeps two-dimensional latent dynamics and adds the VAE. Its advantage emerges most clearly in the eight-dimensional case.
Equation 17
| (17) |
A.6.6 Joint space dynamics
Seven-dimensional joint trajectories provide a Euclidean comparison compatible with all baselines. NCDS has the best reported reconstruction score; accompanying robot reproductions are simulated. The retained literature table distinguishes prior contraction methods and their comparison baselines.
| Paper title | Abbr. | Compared against | Year |
|---|---|---|---|
| Safe Control with Learned Certificates: A Survey of Neural Lyapunov, Barrier, and Contraction methods | 2022 | ||
| Learning Contraction Policies from Offline Data | MPC (ILQR), RL (CQL) | 2022 | |
| Neural Contraction Metrics for Robust Estimation and Control: A Convex Optimization Approach | NCM | CV-STEM, LQR | 2021 |
| Learning Stabilizable Nonlinear Dynamics with Contraction-Based Regularization | CCM-R | 2021 | |
| Learning Certified Control Using Contraction Metric | C3M | SoS (Sum-of-Squares programming), MPC, RL (PPO) | 2020 |
| Learning position and orientation dynamics from demonstrations via contraction analysis | CDSP | SEDS, CLF-DM (Control Lyapunov Function-based Dynamic Movements), Tau-SEDS, NIVF (neurally imprinted vector fields) | 2019 |
| Learning Stable Dynamical Systems using Contraction Theory | C-GMR | SEDS | 2017 |
| Learning Contracting Vector Fields For Stable Imitation Learning | CVF | DMP, SEDS and CLF-DM | 2017 |
| Learning Partially Contracting Dynamical Systems from Demonstrations | CDSP (only position) | DMP, CLF-DM | 2017 |