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.
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.
- 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.
- 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.
- 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 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