From LIMBS
Visual control
Overview:
|
Vision-based control, also known as visual servoing, is a field of robotics in which a computer controls a robot's motion under visual guidance, much like people do in everyday life when reaching for objects or walking through a cluttered room.
Vision-Based Control using Navigation Functions
We seek vision-based controllers (i.e. visual servoing algorithms) that guarantee global, dynamical convergence while maintaining full view of all feature points. Visual servoing algorithms move a robot to register an image-plane feature constellation with a previously stored view. Such a control scheme may be crippled if a feature leaves the camera field-of-view (FOV) or becomes occluded by an object in the scene. Prior contributors have offered novel heuristic solutions to overcome some of these problems, though few offer mathematical guarantees.
To solve the problem we carefully explore the domain of visible configurations of a rigid body, those for which a set of features lie within the FOV and are not occluded by the rigid body itself. The visible domain for many setups, though geometrically complicated in task-space, projects to image plane coordinates as a simple geometrical shape, wherein the visibility constraints appear as the boundary. This observation often enables the construction of a dynamical controller formed of three ingredients:
a simple model space for the “visible set”; a special class of artificial potential functions for navigating the model space; a smooth change of coordinates from the model space to the image plane.
These three ingredients generate a high performance, globally convergent feedback controller, which is guaranteed to respect visibility obstacles. The first movie shows a 6DOF robot holding a rigid body with four coplanar feature points. The robot is visually controlled to align the features points with those from a previously stored 'goal image'. The second video shows the 3DOF Buehgler robot, depicted below, which is viewed with a high-speed computer vision system. By encoding the edge of the FOV as the maximum of a specially designed artificial potential function, high performance, global convergence is achieved
Link: Experimental implementation and videos
Kernel-Based Visual Servoing
Using spatial sampling kernels, a notion borrowed from information theory, our kernel-based visual servoing research aims to provide a unified framework for observing and tracking visual motion. In this way, the research directly connects Dr. Cowan's geometric control work on visual servoing and the Dr. Hager's work on kernel-based visual tracking. The ideas explored promise to spawn a new framework for the design and analysis of more general information-rich sensor-based control systems.
Historically, vision-based control system design is divided into two pieces: a visual processing front end that tracks feature motions on the image plane, and a control system that works to drive those feature trajectories to some desired constellation. Each feature is treated as just a geometric measurement: a point on the image plane. In this setting, "vision" amounts to a source of a geometric measurements, an appealing abstraction because it places the sensor measurement in purely geometric terms.
This tenuous linkage between visual tracking and visual servoing has been the dominant paradigm for the last 30 years, for good reason. It provides appealing modularity, enabling control theorists to tackle visual servoing problems seemingly independent of the environment, offloading "domain specific" tracking problems to vision experts. However, the classical division between vision and control, although convenient, may be ill equipped for the reality of a complex, unstructured world, for at least two reasons.
First, by decoupling vision and control, the vision design problem includes little or no direct information related to the underlying control task that it serves, rendering it difficult or impossible to make intelligent choices as to what to observe, or how to observe it. Conversely, control cannot adapt to a changing visual environment. Thus, neither the vision nor controller design is tuned (much less optimized) for the properties of its counterpart. Consequently, visual tracking algorithms tend to be hand-tailored (often along with the environment) to provide adequate information needed for a specific control algorithm. Second, and perhaps more importantly, the information content of the visual signal may vary considerably during the execution of a control task, and the information needs of the control task may also change depending on the state of world. This makes it critical to adapt the visual tracker, observer, and controller as the task proceeds. For example, a reaching motion for a humanoid robot may start with a large-scale ballistic motion, progress through an increasingly refined sequence of "skills", and ending with a fine, delicate finger motion. Over this large dynamic range, both vision and control may need to adapt. Indeed, there is informal evidence that such adaptive systems can be incredibly robust. Thus, a key problem is to "match" the controller and observer pair to an appropriately designed tracking system so that the two subsystems work synergistically and appropriately for the task at hand.
Images: The American Robot Merlin robot, used for kernel based visual servoing experiments.
