Imagine you're a delivery driver who's only ever navigated on flat grid-streets. Now someone drops you in a medieval hill town — narrow alleys, staircases, curved walls. Your old GPS (Euclidean coordinates) keeps telling you to drive through buildings. What you need is a map that understands the actual surface you're moving on: where the slopes are, how to measure 'straight ahead' when the ground bends under your feet. That's the core problem this paper solves for robots. The committed claim: the authors build a single discrete Riemannian toolkit — logarithmic maps, exponential maps, parallel transport, ambient-space projections — that works directly on polyhedral surface meshes (the triangle soups that CAD and 3D scanners actually produce), and they show it powers three different robot-learning paradigms without requiring smooth-surface assumptions. This isn't the first paper to put robot motion on manifolds, but it's arguably the first to unify discrete mesh geometry with multiple learning families in a single, consistent framework validated on real hardware. The architecture sits squarely in the discrete differential geometry family. Rather than approximating a smooth Riemannian manifold (as prior work using spectral methods or atlas-based parameterizations does), the paper defines its core operators — log/exp maps, geodesic distances, parallel transport — directly on the combinatorial structure of the mesh: faces, edges, vertices, and the transitions between them. The key structural choice is a 'fixed tangent cone' representation at vertices that avoids the singularities plaguing earlier discrete approaches. Computationally, the heaviest lift is geodesic distance (Dijkstra + unfolding), which stays polynomial in face count. Three learning paradigms get instantiated. Dynamic Movement Primitives (DMPs) gain an improved exponential-map computation and forcing-term encoding via parallel transport, yielding better cross-surface generalization and stability. Gaussian Processes get a geodesic kernel with practical positive-definiteness control, enabling regression at arbitrary mesh locations. Riemannian Flow Matching gets mesh-native vector fields that the authors report outperform spectral baselines in generative quality while cutting training time. Each paradigm is benchmarked against its own prior-art method on the same meshes. Integrity is mixed. The simulation comparisons look solid — DMPs are tested against the mesh-based approach of Beik-Mohammadi et al. (2024), GPs against a Euclidean baseline, and RFM against spectral flow matching. But the real-robot experiments, while convincing as existence proofs, are limited: one involves generalizing user-drawn trajectories across different surfaces, the other plans polishing motions on RGB-D reconstructions. There are no independent replications, no pre-registered benchmarks, and numerical superiority claims (e.g., 'better cross-surface generalisation') lean on qualitative or small-sample comparisons rather than large-scale statistical tests. The paper is honest about limitations but doesn't stress-test at scale. The milestone question is about mesh complexity and task diversity. The demonstrations use relatively simple meshes (a few thousand faces) and constrained tasks. The unlock everyone in surface-aware robotics wants is robust learning on meshes with 100K+ faces — full car bodies, architectural surfaces, human anatomy — with real-time replanning. The gap is probably 2-4 years, gated less by theory (which this paper largely solves) and more by geodesic computation speed and integration with real-time perception pipelines. The obvious experiment not run: scaling to significantly larger meshes (50K-200K faces) and benchmarking computational cost against approximate methods like heat-kernel embeddings or neural surface representations (e.g., neural implicit surfaces). The likely reason is compute budget — geodesic computation on large meshes is expensive, and the paper's contribution is the theoretical framework, not a scaling story. Expect the scaling paper as a follow-up.
