Unified Rigid Motion Modeling with Dual Quaternions

36 concepts23 papers1 proposal

In robotics, computer vision, and aerospace engineering, tracking or controlling an object requires representing its position and orientation in 3D space—a combination known as a 6-Degree-of-Freedom (6-DoF) pose. Mathematically, this lives on the Special Euclidean group $SE(3)$.

The traditional way to handle $SE(3)$ is to decouple it: representing rotation using a unit quaternion or a $3 \times 3$ matrix, and translation using a 3D vector. While intuitive, this separation introduces severe mathematical and practical challenges. Dual quaternions solve these issues by unifying rotation and translation into a single, cohesive algebraic structure.

The Problem with Decoupled Poses

When orientation and position are treated as independent variables, algorithms suffer from three core pathologies:

  1. Error Propagation: In calibration or registration tasks, solving for rotation first and then using that result to solve for translation causes any rotational error to cascade and amplify across the translational calculation.
  2. Scale Imbalance: In optimization and deep learning, mixing angular units (radians or degrees) with linear units (meters or centimeters) creates a scale mismatch. Loss functions require arbitrary hyperparameter tuning to balance these disparate units, leading to unstable training or optimization.
  3. Geometric Inconsistency: Decoupled representations ignore the physical reality that rotation and translation are coupled during rigid motion. For example, when a camera moves along a curved trajectory, its orientation changes in tandem with its position. Filtering or interpolating these components separately breaks this geometric relationship, leading to unnatural distortions or poor outlier detection.

The Intuition of Dual Quaternions

Dual quaternions extend classical quaternions by introducing the dual number unit $\epsilon$, where $\epsilon^2 = 0$ (but $\epsilon \neq 0$). A dual quaternion is expressed as:

$$\hat{q} = r + \epsilon d$$

where $r$ is a real quaternion capturing the rotation, and $d$ is a dual quaternion capturing the translation.

The core intuition behind why this works is Chasles' Theorem, which states that any rigid body displacement can be produced by a single screw motion: a simultaneous rotation around an axis and a translation along that same axis.

By binding rotation and translation together via the dual unit, dual quaternions represent this screw motion as a single mathematical entity. The algebraic operations on dual quaternions naturally preserve the geometric coupling between orientation and position.

Eliminating Error Propagation and Scale Imbalance

By optimizing rotation and translation simultaneously rather than sequentially, dual quaternions eliminate cascading errors. For example, in hand-eye and robot-world calibration, Li et al. and Dekel et al. demonstrated that formulating calibration equations within a unified dual quaternion framework prevents error propagation and allows for globally optimal least-squares solutions via simple optimization techniques. Similarly, Phong et al. showed that dual quaternions transform single-view object pose estimation into well-behaved quadratic error functions that can be solved efficiently.

In deep learning, regressing poses directly from data often causes optimization imbalances because networks struggle to weight angular versus linear errors. Schneider et al. bypassed this scale imbalance in multimodal sensor registration by regressing a single dual quaternion, leading to significantly more stable training. Horn et al. leveraged this 8-element vector representation in point cloud registration to construct a geometrically consistent loss function, avoiding the optimization friction common to decoupled predictions.

Preserving Geometric Coupling in Dynamics and Filtering

When tracking moving systems, treating position and orientation as independent entities can lead to physically impossible state estimates. Gilitschenski et al. addressed this by designing a stochastic filter that uses unit dual quaternions to capture the non-linear coupling inherent in rigid-body motions, avoiding the linearization errors of standard filters.

This joint representation is equally powerful for smoothing noisy trajectories. The authors of a 2017 study performed camera pose filtering directly on the Riemannian manifold of unit dual quaternions. Because the representation enforces strict geometric constraints between orientation and position, their framework detects and rejects tracking outliers that would otherwise appear plausible if evaluated independently.

Decoupled Representation:
[ Rotation (Quaternion) ] -----\
                                >-- Processed independently (Ignores physical coupling)
[ Translation (Vector)   ] -----/

Dual Quaternion Representation:
[ Rotation + ε(Translation) ] ---> Processed as a single Screw Motion (Preserves physical coupling)

In control systems, this holistic formulation simplifies complex maneuvers. Sun et al. applied dual quaternions to coordinate spacecraft orbit (translation) and attitude (rotation) simultaneously, resolving the "unwinding" problem and simplifying coupled thruster allocation. For multi-body systems, such as robotic arms mounted on spacecraft, the authors of a 2018 framework showed that dual quaternion algebra unifies rotational and translational kinematics into a single vector space, allowing a single set of Newton-Euler equations to handle various joint types seamlessly.

Avoiding Singularities and Distortions

Representations like Euler angles suffer from gimbal lock, while standard transformation matrices can introduce shearing and scaling distortions when interpolated or averaged. Unit dual quaternions are completely free of singularities and provide a smooth space for interpolation.

Torsello et al. exploited this property for multiview 3D scan registration, using dual quaternions to efficiently average and diffuse poses across a graph structure. In structural mechanics, Sato Martín de Almagro et al. utilized unit dual quaternions to model large deformations in flexible beams without encountering mathematical singularities.

This distortion-free mapping is also vital for modern neural rendering. Liu et al. integrated dual quaternions into 3D Gaussian Splatting to animate complex articulated objects. By using dual quaternions to skin and transform 3D primitives from observed states back to a canonical frame, they avoided the geometric artifacts and distortions typical of standard rotation-translation matrices.

Structural Biases for Machine Learning

Neural networks operate natively in flat Euclidean spaces ($\mathbb{R}^n$), making it difficult for them to respect the curved topology of $SE(3)$. Li et al. addressed this "topology problem" in absolute pose regression by designing a network that natively outputs unit dual quaternions, forcing the model to learn the inherent coupling of 3D space and yielding superior orientation accuracy.

Dual quaternions also provide powerful inductive biases for reinforcement learning and geometric deep learning:

  • Reward Shaping: In high-DoF manipulation tasks like robot-to-robot object handover, Frau-Alfaro et al. designed an RL reward function based on dual quaternions. This unified metric minimized distance and orientation errors simultaneously, smoothing learning trajectories.
  • Equivariance: Grassucci et al. developed a neural network framework that achieves simultaneous rotational and translational equivariance. Because the network processes rigid motions as unified dual quaternions, it generalizes learned motion patterns to any arbitrary starting position or orientation in space, reducing parameter counts by up to 88%.
  • Relational Embeddings: Beyond physical space, Cao et al. extended this geometry to Knowledge Graph Embeddings (DualE). By modeling entities and relations as dual quaternions, their framework captures complex abstract patterns like hierarchy, symmetry, and composition within a single geometric structure via screw motions.

Key Foundational Work

Simultaneous robot-world and hand-eye calibration using dual-quaternions and Kronecker product (2010) — contributed 1 concept elements to this direction.
Optimal estimation of object pose from a single perspective view (1993) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 22 papers
  2. Version 1 · 2026-10-04 · 34 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_dual_quaternion_representation_617edbb7,
  title = {dual quaternion representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-quaternion-representation/rdr_617edbb79bb945c99bd1bf4e73e1717c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_dual_quaternion_representation_617edbb7_v2,
  title = {dual quaternion representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-quaternion-representation/rdr_617edbb79bb945c99bd1bf4e73e1717c/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake