When planning motion for a high-dimensional robot—such as a humanoid arm or a multi-joint industrial manipulator—traditional robotics separates the problem into two distinct steps: kinematics (finding a path through space) and dynamics (calculating the forces and torques required to move along that path). This separation often leads to a fundamental mismatch. A path that looks short and elegant in pure geometric space might require massive, inefficient torques because it fights the robot's natural physical properties, like its mass distribution and joint inertia.
A powerful alternative is to merge geometry and physics into a single mathematical framework using Riemannian geometry. Instead of treating the robot's configuration space as a flat, linear Euclidean space, we can model it as a curved, non-Euclidean manifold. By carefully choosing how we measure "distance" on this manifold, we can ensure that the mathematically shortest paths (called geodesics) automatically correspond to the most physically efficient, obstacle-free trajectories.
The Intuition of a Metric Tensor
In standard Euclidean space, the distance $ds$ between two nearby points is given by the standard Pythagorean theorem: $ds^2 = dx^2 + dy^2$. Every direction is treated equally, and the environment is uniform.
A Riemannian manifold introduces a metric tensor, usually denoted as $M(q)$, which acts as a local modifier to this distance formula. At any configuration $q$, the metric tensor warps space, stretching it in some directions and compressing it in others:
$$ds^2 = dq^T M(q) dq$$
If $M(q)$ is large in a certain direction, moving that way incurs a high "distance" cost. If we then look for a geodesic—the path that minimizes the total integrated distance between two points—the path will naturally bend away from regions where $M(q)$ is large and exploit regions where $M(q)$ is small.
By changing the shape of $M(q)$, we can encode physical effort, obstacles, and task constraints directly into the fabric of the space itself.
Grounding Geometry in Physics: The Kinetic Energy Metric
The most elegant choice for this metric tensor is the robot's own mass-inertia matrix. As argued by Jaquier et al., when the metric tensor $M(q)$ is defined as the system's mass-inertia matrix, the geometric distance along a path corresponds exactly to the square root of the kinetic energy required to execute that motion.
Under this formulation, the physical structure of the robot defines a curved configuration space where geometric distances directly correspond to physical effort. Finding a geodesic on this manifold is mathematically equivalent to finding a trajectory that minimizes kinetic energy.
This principle is not just useful for robots; it appears to be how biology solves motor control. Studies by Liebermann et al. demonstrate that human three-dimensional pointing movements can be modeled by treating the arm's inertial properties as a curved geometric landscape. Similarly, Klein et al. showed that complex human arm movements are planned as sequences of "geodesic synergies" that minimize energy across this inertial landscape. When retargeting human motions to humanoid robots, mapping the movement through this shared inertial geometry preserves the natural, efficient grace of the original motion.
Warping Space to Evade Obstacles
We can extend this geometric intuition to handle environmental obstacles. Instead of treating obstacles as hard logical constraints (e.g., "if distance < threshold, stop"), we can mathematically warp the metric tensor so that obstacles become impassable geometric mountains.
As demonstrated by Mainprice et al., by redefining the metric tensor to grow infinitely large near the boundaries of physical obstacles, the connectivity of the workspace is baked directly into the manifold. An optimizer trying to find a short path through a narrow passage does not need to guess its way through a logical maze; it simply follows the natural "geodesic flow" dictated by the warped space.
Flat Euclidean Space Warped Riemannian Manifold
(Obstacle as a hard boundary) (Obstacle as a geometric hill)
[Start] [Start]
| / \
| / \
[Obstacle] / \
| | (H) | <- Path naturally
| \ / bends around
[Goal] \ / the high cost
[Goal]
For autonomous navigation, Meng et al. showed that a neural network can be trained to predict these local Riemannian metrics and acceleration policies based on visual inputs. By distributing these local metrics across control points on a vehicle's geometry, the system can analytically solve for optimal control commands that respect physical boundaries and vehicle dynamics simultaneously.
Compositionality via Riemannian Motion Policies
In real-world robotics, a system rarely has only one goal. A robot arm might need to move its hand to a target, keep its elbow low to avoid blocking a camera, and keep its entire body away from an obstacle. Each of these tasks lives in a different "task space."
How do we combine these competing geometric demands without them canceling each other out? This is the problem solved by Riemannian Motion Policies (RMPs).
As formulated by Cheng et al., an RMP represents a reactive motion policy as a combination of an acceleration policy and a directional Riemannian metric tensor. The metric tensor acts as a directional weight, expressing how important that specific task space is along different directions of motion.
Using a tree-structured computational graph, RMPs can be mapped from various local task spaces back to the robot's configuration space using a mathematical operation called a pullback. If a task space has a metric $M_x$ and is connected to the configuration space via a Jacobian matrix $J$, the pulled-back metric in configuration space is:
$$M_q = J^T M_x J$$
By summing up the pulled-back metrics and forces from all tasks, the robot can analytically compute a single, globally coordinated joint acceleration that optimally balances every constraint.
Learning Manifolds from Data
We do not always have an analytical formula for the ideal manifold, especially when trying to mimic complex human manipulation skills. In these cases, the Riemannian manifold can be learned directly from human demonstrations.
The authors of Beik-Mohammadi et al. showed that a Variational Autoencoder (VAE) can learn a continuous latent space from human demonstrations, where the data density defines a Riemannian metric. On this learned surface, natural-looking motions are simply the geodesics. Because the policy is embedded in the geometry of the space, the robot can adapt in real time to moving targets or new obstacles without needing to be retrained on specific trajectories.
To make this learning robust when dealing with complex visual inputs, Ding et al. introduced Riemannian Flow Matching Policies (RFMP). This approach ensures that the generated actions strictly adhere to non-Euclidean geometric constraints—such as end-effector orientations that must live on the surface of a sphere—achieving faster training and more stable control than standard flat-space diffusion models.
A Unified Framework for Non-Euclidean Constraints
By moving away from flat-space assumptions, this geometric paradigm provides a mathematically rigorous toolkit for a wide variety of complex robotic problems:
- Orientation and Rotation: End-effector trajectories inherently involve non-Euclidean spaces like the rotation group $SO(3)$. Rozo et al. leveraged Riemannian operations like logarithmic and exponential maps to generate smooth, singularity-free 3D trajectories that outperform traditional Euclidean approximations.
- Imitation and Uncertainty: When learning skills, robots must understand not just the average path, but the allowable variance. Batzianoulis et al. applied structured prediction to learn trajectories and their associated uncertainties directly on Riemannian manifolds, ensuring that the predicted distributions remain mathematically valid.
- Dexterity Tracking: A robot's ability to apply forces varies with its posture, a property captured by a manipulability ellipsoid. Jaquier et al. developed tracking controllers that operate directly on the manifold of Symmetric Positive Definite (SPD) matrices, allowing robots to transfer and track optimal dexterity profiles across different kinematic structures.
- Global Motion Planning: For robots with continuous revolute joints or mobile bases, Cohn et al. demonstrated that configuration spaces can be modeled as graphs of geodesically convex sets, enabling globally optimal, collision-free planning in non-Euclidean spaces.
Ultimately, treating motion generation as a problem of Riemannian geometry allows engineers and researchers to stop fighting the non-linearities of physics and environment. By embedding dynamics, obstacles, and task constraints directly into the metric tensor, the physics of the system does the heavy lifting—turning complex control problems into the simple act of following the shortest path.
Key Foundational Work
Neural Autonomous Navigation with Riemannian Motion Policy (2019) — contributed 1 concept elements to this direction.
A Computational Model for Redundant Human Three-Dimensional Pointing Movements: Integration of Independent Spatial and Temporal Motor Plans Simplifies Movement Dynamics (2007) — contributed 1 concept elements to this direction.