Visual Behaviors for Docking

José Santos-Victor, Giulio SandiniCVIU 1997 · DOI

TL;DR

This paper introduces a novel approach to visual servoing for mobile robot docking using only the normal flow of optical information. This method allows robots to dock to surfaces without requiring explicit 3D reconstruction or camera calibration, offering a robust and efficient solution for both onboard and external docking station scenarios.

What Problem Does This Paper Address?

The paper tackles the challenge of enabling mobile robots to accurately and reliably dock with surfaces. Traditional approaches often rely on complex 3D reconstruction of the environment or precise camera calibration, which can be computationally expensive, sensitive to errors, and difficult to implement in real-time. The authors aim to simplify this process by directly using visual cues – specifically, the optical flow – to control the robot's motion, thereby closing the perception-action loop more tightly and efficiently. They address two distinct docking scenarios: "ego-docking," where the robot has its own camera and computing resources, and "eco-docking," where a central docking station with a camera serves multiple robots.

Methods

The core of the paper lies in its use of normal flow for visual control. Normal flow is the component of optical flow that can be reliably estimated due to the aperture problem, which limits the ability to determine the full 2D flow field from a single image point. The authors propose an affine motion model to describe the optical flow field on a planar surface. They then develop a fast and robust estimation procedure for the affine parameters of this flow field, which is designed to be resilient to outliers.

Crucially, these estimated affine parameters are directly used to drive the robot's motion controllers. The control strategy focuses on two main objectives:

  1. Heading control: Aligning the camera's optical axis with the surface normal, ensuring the robot approaches perpendicularly.
  2. Time-to-crash control: Regulating the robot's forward speed to slow down as it approaches the docking surface, preventing collisions.

This approach bypasses the need for explicit 3D scene reconstruction or camera calibration, relying solely on the direct visual measurements from the normal flow. The same control architecture is shown to be applicable to both ego-docking and eco-docking scenarios.

Datasets

Not discussed in the paper. The experiments were conducted using real-world setups with a TRC Labmate mobile platform and a camera.

Key Insights and Ideas

  • Direct visual control via normal flow: The most significant insight is that robust docking can be achieved by directly using the normal flow component of the optical flow, without needing to solve the full optical flow problem or perform 3D reconstruction. This significantly simplifies the computational requirements.
  • Affine model for planar surfaces: The assumption of planar surfaces (like docking walls or floors) allows for an affine model of optical flow, which is computationally tractable and provides a good approximation for the local motion.
  • Robustness to outliers: The proposed estimation procedure for affine parameters incorporates a mechanism to handle noisy or erroneous flow measurements, leading to more reliable control.
  • Unified approach for ego- and eco-docking: The ability to use the same control framework for both onboard and external docking stations demonstrates the generality and flexibility of the proposed visual behavior.
  • "Vision during action" paradigm: The paper strongly advocates for the "vision during action" approach, where perception and action are tightly coupled, with visual information continuously guiding and monitoring the robot's behavior.

Key Results

The paper demonstrates a real-time implementation of the visual docking behaviors. Experiments showed:

  • Robust docking performance: The system successfully achieved perpendicular alignment with the docking surface and controlled approach speed in both ego-docking and eco-docking scenarios.
  • Smooth control: The robot's forward speed and heading direction evolved smoothly during the docking maneuvers, as visualized in trajectory plots.
  • No need for calibration: The system achieved successful docking without requiring prior knowledge of camera intrinsic or extrinsic parameters, or the vehicle's motion parameters.

Limitations and Future Work

  • Planar surface assumption: The current approach relies on the assumption that the docking surface is planar. While this is a reasonable assumption for many docking scenarios, it might not hold for more complex environments.
  • Focus of Expansion (FOE) as a proxy for "time to crash": The control relies on estimating the time to crash, which is derived from the optical flow. Errors in flow estimation or deviations from planar motion could impact this estimation.
  • Future Work: The authors suggest combining different visual behaviors (like centering, wall following, obstacle avoidance, and docking) to accomplish more complex tasks. This would involve designing a "planner" to select the appropriate behavior based on the robot's situation, rather than trying to understand the entire environment. They envision this as a powerful way to break down complex problems into simpler, manageable behaviors.

Summary updated

BibTeX
@article{santosvictor1997visual,
  author = {José Santos-Victor and Giulio Sandini},
  title = {{Visual Behaviors for Docking}},
  journal = {CVIU},
  year = {1997},
  doi = {10.1006/cviu.1997.0528},
}
Figure