From Movement Primitives to Distance Fields to Dynamical Systems

Yiming Li, Sylvain Calinon · 2025

Read From Movement Primitives to Distance Fields to Dynamical Systems by Yiming Li, Sylvain Calinon (2025) 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-2504-09705.txt.

Introduction

Yiming Li and Sylvain Calinon published this robotics paper in 2025. Robots often learn motions by watching demonstrations. One classic format, movement primitives, stores a motion as a timed trajectory: compact, but tied to the clock. The rival format, an autonomous dynamical system, reacts to disturbances on its own but is harder to shape. The authors propose a bridge that converts the first format into the second.

The trick is to rewrite the trajectory as a shape, encoded by a distance field that says how far each point in space sits from the path. Steering downhill on that field reproduces the motion while naturally absorbing pushes and shifts. Building the trajectories from quadratic Bézier curves makes the distance field computable in closed form, with stability guarantees attached. Simulations and real robot trials showed the method encoding complex motions while keeping them stable.

Catalog record

CREATORYiming Li, Sylvain Calinon
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-09705

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

From Movement Primitives to Distance Fields to Dynamical Systems Yiming Li, Sylvain Calinon Abstract Developing autonomous robots capable of learning and reproducing complex motions from demonstrations remains a fundamental challenge in robotics. On the one hand, movement primitives (MPs) provide a compact and modular representation of continuous trajectories. On the other hand, autonomous systems provide control policies that are time independent. We propose in this paper a simple and flexible approach that gathers the advantages of both representations by transforming MPs into autonomous systems. The key idea is to transform the explicit representation of a trajectory as an implicit shape encoded as a distance field. This conversion from a time-dependent motion to a spatial representation enables the definition of an autonomous dynamical system with modular reactions to perturbation. Asymptotic stability guarantees are provided by using Bernstein basis functions in the MPs, representing trajectories as concatenated quadratic Bézier curves, which provide an analytical method for computing distance fields. This approach bridges conventional MPs with distance fields, ensuring smooth and precise motion encoding, while maintaining a continuous spatial representation. By simply leveraging the analytic gradients of the curve and its distance field, a stable dynamical system can be computed to reproduce the demonstrated trajectories while handling perturbations, without requiring a model of the dynamical system to be estimated. Numerical simulations and real-world robotic experiments validate our method’s ability to encode complex motion patterns while ensuring trajectory stability, together with the flexibility of designing the desired reaction to perturbations. An interactive project page demonstrating our approach is available at https://idiap.github.io/mp-df-ds/ . I Introduction and Related Work Learning complex motion skills by optimization or through demonstrations remains a fundamental challenge in robotics, which can computationally be addressed from different perspectives. Movement primitives (MPs) offer a compact and modular framework for encoding, generalizing, and reproducing demonstrated trajectories by superposition of basis functions, see [ 1 ] for a review. Various MP formulations have been proposed, including dynamical movement primitives (DMPs) [ 2 , 3 ] , Gaussian mixture regression (GMR) [ 4 ] , probabilistic movement primitives (ProMPs) [ 5 ] , kernelized movement primitives (KMPs) [ 6 ] , or Fourier movement primitives (FMPs) [ 7 ] . Conventional MPs are often limited in their ability to robustly handle perturbations and dynamically adapt to environmental changes, due to their explicit representation of the signals based on a time or phase variable. This downside has motivated the development of alternative autonomous systems formulations to provide time-independent dynamical system representations, inherently robust to perturbations, at the expense of complexifying the problem of modeling long and/or high-dimensional movements. Examples include SEDS [ 8 ] , CLF-DM [ 9 ] , Imitation Flow [ 10 ] , and Neural Contractive Dynamical Systems (NCDS) [ 11 ] , which all ensure stability and adaptability by learning or imposing constraints on the system dynamics. Fig. 1 : From movement primitive ( left ) to distance field ( middle ) to dynamical system ( right ). Left: A S-shape trajectory is parameterized using the concatenation of 2 quadratic curves. Control points are shown in blue. Middle: Distance field of the quadratic spline. The gradient of the distance field at a given point (in red) is perpendicular to the curve at its closest point (in orange). This property is used to compute the distance field analytically and build the stable dynamical system ( Right ). Red and black bullets are a random start point and the equilibrium point at the end of the trajectory. We propose in this paper a simple and flexible approach that gathers the advantages of both representations by transforming MPs into autonomous systems. Our work follows the same foundational principle of using autonomous systems for robust motion generation, but differs by avoiding learning-based modeling altogether. Instead, we directly leverage the geometric properties of a reference trajectory encoded as a distance field, allowing us to define an autonomous system with asymptotic stability guarantees through analytic computation. A related concept has been explored in contact-aware tasks such as robotic polishing, where surfaces and normal vectors guide the design of dynamical systems [ 12 ] . We apply the idea to full trajectory encoding beyond surface interaction. The core idea is to reinterpret the demonstrated trajectories as implicit spatial representations encoded as distance fields. Rather than representing motion as a sequence of time-parameterized states, we encode the shape of the trajectory in space. The resulting field defines the shortest distance to the reference trajectory at each spatial location, enabling motion generation as gradient descent over this field. Asymptotic stability is ensured by incorporating Bernstein basis functions into movement primitives (MPs), representing trajectories as concatenated quadratic Bézier curves. This formulation enables the analytical computation of distance fields, eliminating the need for point sampling or heuristic weighting [ 13 ] . By bridging conventional MPs with distance fields, this approach provides smooth and precise motion encoding within a continuous spatial representation. Leveraging the analytical gradients of both the curve and its distance field, a stable dynamical system can be directly derived to reproduce demonstrated trajectories and handle perturbations, without requiring explicit system modeling or learning techniques. The motivation of representing structure implicitly in continuous space is inspired by recent advances in Signed Distance Fields (SDFs) and Neural Radiance Fields (NeRFs) [ 14 , 15 ] , offering compact and continuous representations for shapes and scenes. These techniques have influenced robotics, particularly in motion planning and scene understanding [ 16 , 17 ] . Our work in this paper instead focuses on representing 1D trajectories embedded in high-dimensional spaces. By exploiting this lower-dimensional structure, we derive closed-form distance fields from concatenated quadratic Bézier curves, enabling smooth, precise, and efficient computation of spatial gradients for motion generation. This spatial formulation offers several advantages. It eliminates the need for time synchronization, naturally supports modular responses to perturbations, and provides a continuous and robust encoding of motion by avoiding the discretization artifacts common in numerical methods. The main contributions of this paper are: • We propose a novel use of distance fields as an implicit trajectory representation for autonomous systems, shifting from conventional time-dependent trajectory encoding to a spatial formulation guided by distance gradients. • We derive an analytical method for computing distance fields to reference trajectories represented by concatenated quadratic Bézier curves. This approach enables efficient and accurate computation of distance and gradient fields while preserving the ability to represent complex motion patterns. • We show that combining movement primitives with their corresponding gradient fields naturally yields autonomous dynamical systems with provable asymptotic stability—without requiring system modeling or learning-based estimation. We validate our approach through numerical simulations and real-world robot experiments, demonstrating its effectiveness in encoding complex motions, generating stable and smooth trajectories, and adapting to external perturbations. Additionally, we highlight the simplicity and flexibility of our framework in: • integrating multiple demonstrations through simple field operations; • scaling to high-dimensional settings such as robot joint spaces; • enabling adaptive motions that achieve a balance between robustness to perturbations and generalization to new trajectories. II Representing Trajectories as Distance Fields Given a reference trajectory \mathbf{F} whose points are given by \mathbf{f}(t):[0,T]\to\mathbb{R}^{D} parameterized by t\in[0,T] , where T is the trajectory length and D is the spatial

Read the complete text