Technical report on Optimization-Based Bearing-Only Visual Homing with Applications to a 2-D Unicycle Model

Roberto Tron, Kostas DaniilidisarXiv 2014 · DOI · arXiv

Abstract

This paper addresses the challenge of bearing-based visual homing, where a mobile robot must return to a "home" location using only the angular directions (bearings) of known landmarks. We propose a control framework based on the gradient field of a Lyapunov cost function and provide formal proofs for global convergence. Notably, we show that the widely used Average Landmark Vector (ALV) method is a specific instance of our framework. We further extend this logic to non-holonomic unicycle models using sliding mode control. Finally, we characterize the sensitivity of the home location to measurement noise, providing insights into optimal landmark configurations.

Problem Definition: Homing without Range

Visual homing involves guiding a robot back to a previously visited location based on visual cues. In many scenarios, a robot's camera can accurately determine the bearing to a landmark but cannot easily estimate its distance. This "bearing-only" constraint makes the problem geometrically challenging.

Geometry of the bearing-only visual homing problem, showing the relationship between the current location (x), home location (xg), and landmarks (xi).

As shown in the figure above, the robot at position $x$ observes landmarks at $x_i$ with bearings $\beta_i$. The goal is to reach $x_g$, where the observed bearings would match the target bearings $\beta_{gi}$. Previous heuristic methods, like the Average Landmark Vector (ALV), lacked formal proofs of global convergence, leaving open the possibility of the robot getting stuck in local minima.

Methodology: Gradient-Based Control

We frame homing as an optimization problem by defining a Lyapunov cost function that penalizes the difference between current and target bearings.

  1. Holonomic Gradient Control: For robots capable of moving in any direction, we use the negative gradient of this cost function as the velocity command. We prove that as long as landmarks are not all collinear with the home position, the cost surface is "well-behaved" with a unique global minimum at the home location.
  2. The ALV Connection: A key contribution is the discovery that the classic ALV method is mathematically equivalent to our gradient descent approach when using a specific "reshaping function." This provides the first formal theoretical validation for why ALV works in practice.
  3. Unicycle Extension: Real-world robots often follow unicycle kinematics (they cannot move sideways). We address this by implementing a Sliding Mode Control law. This controller adjusts the robot's orientation to align its forward motion with the ideal gradient path, ensuring convergence even from unfavorable starting orientations.

Results and Insights

Our simulations confirm that the proposed controllers are robust and globally convergent. The robot successfully reaches the home location from various starting points and orientations.

The cost surface (contour lines) and robot trajectories (black lines) converging to the home location (green dot) amidst landmarks (blue dots).

The figure above visualizes the "cost surface" created by the landmarks. The gradient field naturally pulls the robot toward the global minimum. Our sensitivity analysis further reveals that the stability of the home location depends heavily on landmark geometry. Homing is most robust when the robot is surrounded by landmarks; if landmarks are clustered on one side, small errors in bearing measurements can significantly shift the perceived home position.

Limitations and Future Work

While robust, the current framework assumes the robot has access to a "global compass" to maintain a consistent orientation frame for bearing measurements. Future research will focus on "compass-free" homing using relative bearing changes and epipolar geometry. Additionally, while the gradient framework is generalizable to 3D, the specific unicycle control laws presented here are tailored for 2D ground-based navigation.

Summary updated

Abstract

We consider the problem of bearing-based visual homing: Given a mobile robot which can measure bearing directions with respect to known landmarks, the goal is to guide the robot toward a desired "home" location. We propose a control law based on the gradient field of a Lyapunov function, and give sufficient conditions for global convergence. We show that the well-known Average Landmark Vector method (for which no convergence proof was known) can be obtained as a particular case of our framework. We then derive a sliding mode control law for a unicycle model which follows this gradient field. Both controllers do not depend on range information. Finally, we also show how our framework can be used to characterize the sensitivity of a home location with respect to noise in the specified bearings. This is an extended version of the conference paper [1].
BibTeX
@misc{tron2014technical,
  author = {Roberto Tron and Kostas Daniilidis},
  title = {{Technical report on Optimization-Based Bearing-Only Visual Homing with Applications to a 2-D Unicycle Model}},
  year = {2014},
  eprint = {1402.3584},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1402.3584},
}
Figure