Abstract
In robotics, data often resides on non-Euclidean manifolds, such as the space of rotations or positive-definite matrices. A prevalent but mathematically flawed shortcut involves projecting this data onto a single tangent space to utilize standard Euclidean machine learning algorithms. This paper formalizes this "single tangent space fallacy," detailing the resulting geometric distortions and model failures. By contrasting this approach with intrinsic Riemannian methods—which leverage parallel transport and multiple tangent spaces—we demonstrate significant performance gains in density estimation and dynamical system learning across various high-dimensional manifolds.
Introduction
Modern robotics frequently encounters data constrained to curved surfaces, known as Riemannian manifolds. Examples include unit quaternions for orientations ($S^3$), joint configurations of redundant manipulators, and symmetric positive-definite (SPD) matrices for stiffness or inertia. While standard machine learning assumes a flat Euclidean geometry, applying these tools directly to manifold-valued data leads to suboptimal or physically nonsensical results.
To bridge this gap, many researchers employ a "tangent space approximation": they select a base point, project all data onto the tangent plane at that point via the logarithmic map, perform learning in this flat space, and project back via the exponential map. This paper argues that this shortcut is a fundamental misunderstanding of differential geometry that fails as data variance or dimensionality increases.
The Single Tangent Space Fallacy
The "fallacy" stems from treating a local linear approximation as a global one. Several critical issues arise:
- Locality and Injectivity: The exponential map is only a local diffeomorphism. Data extending beyond the manifold's injectivity radius results in overlapping projections and "leaking" probability densities.
- Distance Distortion: Tangent spaces only preserve distances relative to the base point. Distances between other points are systematically distorted, undermining kernel-based and distance-based learning.
- Physical Incoherence: Treating differences between projected points as velocities is mathematically incorrect. True velocities are vectors attached to specific points on the manifold; treating them as independent entities in a single flat plane leads to distorted trajectories in dynamical systems.
Methods: Intrinsic Riemannian Learning
The authors advocate for algorithms formulated directly on the manifold. Unlike the single tangent space approach, intrinsic Riemannian methods utilize multiple tangent spaces and parallel transport—a mechanism to move vectors between tangent spaces while preserving their geometric properties.
The paper evaluates three paradigms:
- Euclidean: Ignoring the manifold structure entirely.
- Tangent Space Approximation: The flawed single-projection method.
- Riemannian: Leveraging the manifold's global structure through coordinate-invariant operations.
These are tested using Gaussian Mixture Models (GMMs) for density estimation and Continuous Normalizing Flows (CNFs) for learning stable dynamical systems.
Experimental Results
The study utilized synthetic benchmarks on spheres ($S^d$) and SPD manifolds ($S^d_{++}$), alongside the LASA handwriting dataset projected onto $S^2$.
Density Estimation
On a 10D sphere ($S^{10}$), the Riemannian GMM achieved a log-likelihood of approximately 1068, significantly outperforming the Tangent GMM's 949. The Tangent GMM consistently failed to represent distributions that wrapped around the manifold, a common occurrence in high-dimensional spaces.
Dynamical Systems
When learning constrained motions from the LASA dataset, the Riemannian approach maintained a 99-100% success rate in reaching target goals. Conversely, the Tangent Space approach dropped to a 67% success rate for complex trajectories (e.g., 'W' shapes), producing distorted paths that failed to respect the demonstrated curvature.
Conclusion
The "single tangent space fallacy" is a pervasive issue in robot learning that appears benign in low-dimensional, low-variance scenarios but fails catastrophically in more complex settings. To build robust and geometrically consistent robot learning systems, researchers must move toward intrinsic Riemannian methods. Future work should focus on developing coordinate-invariant reinforcement learning and extending these principles to manifolds with non-constant curvature.
Summary updated