Curve-Induced Dynamical Systems on Riemannian Manifolds and Lie Groups
Saray Bakker, Martin Schonger, Tobias Löw, Javier Alonso-Mora, Sylvain Calinon · 2026
Read Curve-Induced Dynamical Systems on Riemannian Manifolds and Lie Groups by Saray Bakker, Martin Schonger, Tobias Löw, Javier Alonso-Mora, Sylvain Calinon (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-05268.txt.
Introduction
Saray Bakker and team present CDSM, a real-time method building dynamical systems on Riemannian manifolds and Lie groups for shape-respecting robot actions. It builds a nominal curve on the manifold, mixing tangential motion along the curve with normal attraction toward it. Balance analysis follows, with an S2 benchmark beating leading methods on path accuracy, path drift, and making and query times. Manipulator tests adapting SE(3) poses and damping matrices online, plus a mobile manipulator, show real use.
Catalog record
| CREATOR | Saray Bakker, Martin Schonger, Tobias Löw, Javier Alonso-Mora, Sylvain Calinon |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 69869 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.05268
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Curve-Induced Dynamical Systems on Riemannian Manifolds and Lie Groups Saray Bakker, Martin Schonger, Tobias Löw, Javier Alonso-Mora, Sylvain Calinon Abstract Deploying robots in household environments requires safe, adaptable, and interpretable behaviors that respect the geometric structure of tasks. Often represented on Lie groups and Riemannian manifolds, this includes poses on S E ( 3 ) SE(3) or symmetric positive definite matrices encoding stiffness or damping matrices. In this context, dynamical system–based approaches offer a natural framework for generating such behavior, providing stability and convergence while remaining responsive to changes in the environment. We introduce dsmpman ( dsmpman ), a real-time framework for constructing dynamical systems directly on Riemannian manifolds and Lie groups. The proposed approach constructs a nominal curve on the manifold, and generates a dynamical system which combines a tangential component that drives motion along the curve and a normal component that attracts the state toward the curve. We provide a stability analysis of the resulting dynamical system and validate the method quantitatively. On an S 2 S^{2} benchmark, dsmpman demonstrates improved trajectory accuracy, reduced path deviation, and faster generation and query times compared to state-of-the-art methods. Finally, we demonstrate the practical applicability of the framework on both a robotic manipulator, where poses on S E ( 3 ) SE(3) and damping matrices on 𝒮 + + n \mathcal{S}_{++}^{n} are adapted online, and a mobile manipulator. I Introduction Real-world deployment of robots in household settings requires systems that are inherently safe, adaptable, and interpretable. Achieving such safe and reliable behavior requires models that respect the underlying structure of the complex tasks these robots perform. In many cases, the processed data is inherently non-Euclidean, residing on curved manifolds rather than in standard vector spaces. Considering the geometry of the task, such as poses on SE(3) , or stiffness and damping expressed as spd ( spd ) matrices on \mathcal{S}_{++}^{n} , is crucial for effective task execution as well as safety during interaction. For instance, in a robot-assisted dressing task, as shown in Fig. 1 , the robot adapts its end-effector pose to the human arm’s configuration and varies the affine transformation for the desired velocity in the impedance controller, hereafter referred to as the damping matrix, throughout the interaction. Dynamical system–based approaches offer an appealing framework for generating robot behavior in such settings, as they can provide formal guarantees on stability and convergence while remaining highly reactive to changes in the environment. Several approaches have been developed, ranging from learning-based to deterministic methods. Fig. 1 : We consider a dressing task in which the robot adapts both its pose and the damping matrix of an impedance controller throughout the motion. External perturbations are applied by physically pushing the robot, and we observe that the proposed dsmpman converges back to the desired nominal trajectory, given the tracked configuration of the human arm via a camera. The illustration in the bottom-left corner was generated using Google Gemini. Recently, there has been growing interest in learning stable ds on Riemannian manifolds and Lie groups. A prominent line of work leverages normalizing flows as diffeomorphic mappings, both for Lie groups [ 1 ] and for more general Riemannian manifolds [ 2 ] . Closely related approaches employ Flow Matching on Riemannian manifolds [ 3 , 4 ] or enforce stability explicitly through constraints in the loss function [ 5 ] . Broadly speaking, data-driven methods still make specific assumptions to achieve stability that could limit performance on manifolds. They typically require several demonstrations, and training times hinder real-time adaptation of the learned DS to changes in the environment or task, such as adapting to different human arm configurations in the dressing task. In addition, convergence in a learned ds is typically directed towards the goal, rather than the demonstrated trajectories, even though explicitly attracting the system to those trajectories could improve user-interpretability and task execution, especially when working with a small dataset. This motivates the use of vector-field–based approaches defining a reference curve and constructing a vector field that draws motion toward this curve while also steering it along the curve’s direction. These curves can be constructed as mps ( mps ), an effective and lightweight approach to encode a trajectory or data sequence, which can be defined on Riemannian manifolds and Lie groups [ 6 , 7 ] . In Euclidean space, Li and Calinon [ 8 ] show that an analytic distance map can be generated when using composite quadratic Bézier curves, and converted into a dynamical system to react to disturbances. Although extensions to Riemannian manifolds are proposed, applying the approach to arbitrary demonstrated trajectories remains nontrivial [ 9 ] , or, relies on projecting actions from \mathbb{R}^{n} onto the tangent spaces of the manifold without explicitly accounting for the geometric structure to construct a dynamical system on the manifold [ 10 ] . We propose dsmpman ( dsmpman ), which consists of a dynamical system formulation on Riemannian manifolds and an extension to Lie groups, naturally handling robotic data such as varying damping matrices, poses, and orientations, while maintaining reactivity and practical stability. Building on the Euclidean-space DS in [ 8 ] , we leverage tools from differential geometry to generate a stable ds converging to a curve for Riemannian manifolds, such as S^{2} and \mathcal{S}_{++}^{n} -matrices, and Lie groups, such as SO(3) and SE(3) . In particular, gain matrices of an impedance controller can be represented as full \mathcal{S}_{++}^{6} -matrices [ 11 ] , enabling extensions beyond approaches that leverage diagonal stiffness and damping matrices [ 12 , 13 ] . The curve is built from single or multiple demonstrated trajectories or data-sequences, thereby ensuring C^{1} -continuity. We provide fast curve optimization as well as batch evaluation and forward simulation of the ds using JAX [ 14 , 15 ] . Moreover, we propose a coupling between a spatial profile, encoding poses, and a curve of \mathcal{S}_{++}^{6} matrices, encoding damping matrices. Our approach is able to compute both the parametrized curves as well as actions of the ds in real-time, making it well-suitable for several real-world tasks such as the dressing task in Fig. 1 , where high responsiveness to changes of the arm configuration and disturbances is required. Additionally, to extend the adaptability of the dsmpman framework, we introduce a phase modulation layer on top of the reference data-based trajectory profile, leveraging the same parametric curve structure. This structure enables flexible adaptation of the system’s timing. For instance, the phase profile can be optimized given velocity constraints or to increase reactivity to dynamic obstacles or agents, without altering the spatial curve itself. By decoupling timing from the curve, dsmpman can adapt its execution speed and responsiveness while preserving the learned motion’s shape. Our contributions are as follows: • A curve-induced dynamical system-formulation suitable for Riemannian manifolds and an extension to Lie groups, which can be constructed from a single or multiple demonstrated trajectories, or data sequences in real-time. • Coupling a dsmpman of poses in SE(3) with a curve of full \mathcal{S}_{++}^{n} matrices encoding variable Cartesian damping matrices by synchronizing their phase profiles. • Combining dsmpman with a phase modulation layer to enable adaptive timing, illustrated by an example imposing speed constraints. •