Traveling the Robot Learning Manifold || Nov 13, 2026

Traveling the Robot Learning Manifold || Nov 13, 2026

Title history
  • Traveling the Robot Learning Manifold || Nov 13, 2026 — observed on 20261113-2254 UTC (YouTube)
🎙 Noémie Jaquier 👥 4K 📅 November 13, 2026 ⏱ 60 min 👁 3 📄 science communication 🧭 2026-11-13
Available in: English (current) Français

Keywords

Riemannian manifoldsflow matching policiesstable dynamical systemsgeometric inductive biasesdeformable object manipulation

Summary

The seminar by Prof. Noémie Jaquier, hosted by the CRUNCH group, presents a comprehensive exploration of how differential geometry and physics-based inductive biases can enhance robot learning. The talk begins with an introduction to the manifold structure of robot state spaces (Euclidean, sphere, torus, SPD matrices, etc.) and demonstrates, via a simple mean computation on a sphere, that ignoring geometry yields incorrect results. Three application areas are then discussed. First, imitation learning with Riemannian flow matching policies, which generalize flow matching to manifolds to encode full poses (position and orientation) and produce dexterous manipulation skills. Second, stable flow matching dynamical systems, which incorporate control-theoretic Lyapunov functions to guarantee convergence to an equilibrium, even under external disturbances, and can scale to high-dimensional state spaces (e.g., 1682-d). Third, symmetric policies via vector-field lifting, which compose multiple symmetries (morphological, rotational, scaling) across different spaces to improve generalizability and data efficiency. Finally, the talk addresses the challenge of deformable objects, proposing a physically-grounded digital twin based on a kinematic chain representation with Lagrangian dynamics, trained from multi-view RGB video using Gaussian splatting. The presentation emphasizes explicit encoding of geometric structures leads to better performance, data efficiency, and safety guarantees.

195 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides high value by unifying several advanced concepts (Riemannian geometry, generative models, control theory, symmetry) into a coherent framework for robot learning. The argumentation is solid: each method is motivated by a clear limitation (e.g., Euclidean averaging on spheres fail), then the geometry-aware solution is introduced with mathematical detail (geodesics, tangent spaces, exponential maps). Demonstrations and videos of real robotics experiments reinforce the claims. The speaker carefully distinguishes between theoretical guarantees and practical implementation (e.g., physics in the controller). The discussion with the audience (Q&A) further clarifies edges cases (plasticity, external disturbances) and justifies design choices. Overall, the value is outstanding for the target audience due to the depth and clarity.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates high scientific rigor: all methods are rooted in established mathematics (differential geometry, Lie groups, flow matching), and each technique is validated with simulations and real-robot experiments. The speaker cites her own and others’ work implicitly, but no external citations are listed in the description. The title accurately reflects the content, as the talk indeed surveys a ‘manifold’ of research directions. The organisation is logical, moving from simple geometric examples to increasingly complex systems. No significant gaps or logical flaws are apparent; limitations are acknowledged (e.g., elastic-only deformable objects). The talk likely relies on published peer-reviewed work, but without explicit sources, we cannot verify connections to the broader literature. No audience comments were provided, so no trends can be inferred.

250 words

Title / Content Match

The title 'Traveling the Robot Learning Manifold' aligns well with the content, which explores how differential geometry and manifold structures guide robot learning algorithms. The metaphor of 'traveling' reflects the journey through geometric spaces.

Quality & Reliability

9/10

The talk is given by a recognized researcher (KTH, EPFL PhD) and presents rigorous mathematical frameworks (differential geometry, group theory) applied to robot learning. The content is well-structured, self-contained, and addresses state-of-the-art techniques. Slide quality and presentation clarity are high, with Q&A clarifying technical details.

Key Moments

Contribution & Novelties

The talk synthesizes recent advances in geometry-aware robot learning, presenting a coherent narrative of how differential geometry serves as an inductive bias across different scales (actions, policies, symmetries, and object dynamics). The specific contributions include a clear demonstration of the failure of Euclidean methods on spheres, a novel integration of flow matching with Riemannian geometry, a stability-preserving extension for dynamical systems, and a unified symmetry composition approach. Additionally, the deformable-object digital twin using Gaussian splatting and Newton-Euler dynamics is a fresh contribution. These ideas together highlight the importance of respecting underlying geometric structures for data-efficient and safe robot learning.

Pour aller plus loin :

  • Riemannian manifold — Provides the mathematical foundation for non-Euclidean geometry, essential to the talk.
  • Flow matching — The generative model base that the speaker generalizes to manifolds (URL might not be stable, but the paper is known). To avoid unstable URLs, one could cite the Wikipedia article on Generative flow networks? Actually, flow matching is not on Wikipedia. Better: Diffusion models which are related but not exact. I’ll use Continuous normalizing flows for the concept of ODE-based generative models.
  • Lie group — Used for orientation and SE(3) in robot learning.
  • Gaussian splatting — The rendering technique used in the deformable object model.
  • Lagrangian mechanics — The physical formalism for the deformable object dynamics.

217 words

Radar Profile

The radar profile shows very high scores in all dimensions, particularly in technical level and information quality, reflecting a deep, well-supported presentation. The only slightly lower (still high) score is in information quantity, possibly because the talk focuses on a few selected examples rather than an exhaustive literature sweep. Overall, the talk is a polished expert exposition.

Reliability 9/10