Technical Summary: Vector Fields for Path Following on Lie Groups with Application in Robot Control

Problem Statement

Many robotic systems, such as omnidirectional aerial vehicles, underwater robots, and manipulator end-effectors, possess independent control over position and orientation (pose). These systems operate as rigid bodies in 3D space, evolving on the configuration manifold $SE(3)$ rather than Euclidean space. The core problem addressed is path following: guiding a system to a geometric curve without a pre-assigned time parametrization. Unlike trajectory tracking, path following decouples spatial tasks from temporal specifications, offering advantages such as improved robustness to disturbances, smoother convergence, and reduced control effort.

While guiding vector fields have been successfully applied to path following in Euclidean space ($\mathbb{R}^n$), extending this framework to Lie groups like $SE(3)$ presents challenges. Previous approaches often embed the manifold in a higher-dimensional Euclidean space and project control inputs back, resulting in redundant control dimensions (e.g., 9 dimensions for $SO(3)$ instead of the intrinsic 3). Furthermore, existing methods on manifolds often rely on implicit level-set definitions of curves, which can be inconvenient for specifying complex paths. This paper addresses the need for a general, intrinsic vector-field framework for path following on connected matrix Lie groups that yields minimal, non-redundant control inputs (e.g., body twists) and supports parametric curve representations.

Methodology

Theoretical Framework

The authors propose a general vector-field strategy for connected matrix Lie groups $G$. The system dynamics are modeled as a first-order, fully-actuated system:
$$\dot{H}(t) = S(\xi(t))H(t)$$
where $H(t) \in G$ is the state, $\xi(t) \in \mathbb{R}^m$ is the control input (generalized twist), and $S$ is a linear map from $\mathbb{R}^m$ to the Lie algebra $\mathfrak{g}$.

The guiding vector field $\Psi(H)$ is constructed as a sum of two components:
$$\Psi(H) = k_N(H)\xi_N(H) + k_T(H)\xi_T(H)$$

1. Normal Component ($\xi_N$): Responsible for convergence to the curve. It is derived from the gradient of a distance function.
2. Tangent Component ($\xi_T$): Responsible for traversal along the curve. It is defined by the tangent vector of the curve at the nearest point.

Key Mathematical Innovations

To extend the Euclidean vector field approach (specifically from [11]) to Lie groups, the authors introduce specific operators and distance properties:

• Operators:
  • $\Xi$ Operator: Extracts the Lie algebra velocity vector from a differentiable path on the group, analogous to the time derivative $\frac{d}{dt}$.
  • $\mathcal{L}$ Operator: Analogous to the gradient operator $\nabla$ on a Lie group, defined via the limit of directional derivatives along the group structure.
• Distance Function Requirements: The framework relies on an Element-to-Element (EE-) distance function $\bar{D}$ that satisfies three critical properties:
  1. Left-Invariance: $\bar{D}(AV, AW) = \bar{D}(V, W)$. This ensures the normal and tangent components are orthogonal.
  2. Chainability: The distance between two points equals the sum of distances along a specific path connecting them. This property guarantees the absence of local minima outside the target curve.
  3. Local Linearity: The distance behaves linearly near the path, ensuring the gradient (normal component) never vanishes outside the curve.

Specialization to $SE(3)$

For the specific case of $SE(3)$, the paper defines a class of Special Exponential Lie Groups. For these groups, a valid EE-distance is constructed using the Frobenius norm of the matrix logarithm:
$$\bar{D}(V, W) = \| \log(V^{-1}W) \|_F$$
This formulation allows for an explicit, efficient algorithm to compute the distance and its derivatives, avoiding the computational expense of general geodesic calculations. The control input $\xi$ corresponds directly to the robot's twist (linear and angular velocities), ensuring the input dimension matches the system's degrees of freedom.

Key Contributions

1. Generalization to Lie Groups: The paper extends the vector field strategy from Euclidean space to connected matrix Lie groups, resulting in a non-redundant control input that respects the intrinsic geometry of the configuration space.
2. Formal Proofs and Distance Properties: The authors provide formal proofs for convergence and traversal, deriving specific properties (left-invariance, chainability, local linearity) that a distance function must satisfy to ensure the vector field's stability.
3. Efficient Algorithm for $SE(3)$: An efficient algorithm is developed for implementing the strategy on $SE(3)$, utilizing the special exponential property to compute distances and gradients without redundant dimensions.
4. Real-World Validation: The approach is validated through experiments with a Kinova Gen3 robotic manipulator tracking complex pose paths.

Experimental Results

The authors conducted experiments using a 7-DOF Kinova Gen3 manipulator. The end-effector was tasked with tracking a complex, discretized curve in $SE(3)$ that avoided self-collisions and joint limits.

• Performance: The system successfully converged to the desired curve and traversed it.
• Metrics: The Euclidean distance (EC-distance) to the curve decreased over time, and the system maintained low position and orientation errors once on the path.
• Computational Efficiency: The vector field computation averaged 8.9 ms per iteration on a standard CPU. The authors note that 99.5% of this time was spent solving the optimization problem to find the nearest point on the curve, suggesting potential for further speedup via parallelization (e.g., GPUs).
• Observations: Minor deviations in error were observed when the robot operated near singular configurations, requiring frequent orientation adjustments, but the system remained stable.

Significance and Claims

The paper claims to offer a unified framework that integrates path planning, path following, and control for robotic systems operating on Lie groups. Its primary significance lies in:

• Intrinsic Control: Unlike methods that project controls from higher-dimensional spaces, this approach generates control inputs (twists) that are minimal and directly applicable to engineering systems.
• Parametric Flexibility: By adopting parametric curve representations rather than level-set intersections, the framework accommodates complex curves more naturally, which is crucial for practical robotic tasks.
• Robustness: The theoretical guarantees of "almost global" convergence (with the ability to escape problematic sets via small perturbations) and the decoupling of spatial and temporal tasks provide practical benefits for real-world navigation in cluttered environments.

The authors conclude that while the current distance metric is specific to special exponential Lie groups, the framework opens avenues for future work on time-varying curves, underactuated systems, and more general distance functions. An open-source implementation is provided to facilitate further research and application.