Design and Calibration of Robotic Systems with Applications in Cooperative Repair and Ultrasound Calibration
Johns Hopkins University
As the need and desire for independently thinking and acting machines increases, current robotic design methods will need to be reevaluated. Robotic agents will work cooperatively in teams and will need to address fault states in the team and the environment, without human intervention, if they are to be considered truly autonomous. Two related aspects of increasing the autonomy and effectiveness of robotic systems are addressed in this work. In the first section, a new test-bed on which principals of cooperative repair and diagnosis for a cooperative multi-agent system (CMS) can be designed and evaluated. The Hexagonal Distributed Modular Robot (HexDMR) evolved over three design iterations. In the Mark 1, a system was designed in which each subsystem (e.g. power, drive mechanisms, the micro-controller) was contained in a module assembled in a hexagonal frame. To repair a subsystem, the ineffective module could be removed and replaced with a new functioning module. In the next iteration of the design, the Mark 2, many aspects of the Mark 1 were improved. The Mark 2 eschewed some of the large, unwieldy and imprecise aspects of the Mark 1 to improve performance. In the Mark 2, each module is attached to it's neighbor, still forming a hexagonal module lattice, but creating a smaller, lighter and more compact system. The Mark 3 design is the final iteration of the HexDMR system. The functionality and homogeneity have been improved in the Mark 3 system by creating a design that draws on the strengths of both the Mark 1 and the Mark 2 iterations. A simpler, more error tolerant module connecting mechanism is used which increases the accuracy and repeatability of the repair maneuvers. In some modules, additional processing circuits have been added and, while this increases the complexity of some of the modules, the distributed control scheme that is now possible greatly simplifies the type and number of connections between the modules. Overall, the new distributed control greatly improves the functionality of the system.Through the improvements that were made over each iteration, the Mark 3 is able to successfully demonstrate autonomous repair maneuvers and serve as a testbed for future cooperative repair and diagnosis research. One of the most important aspects of any robotic system, including the HexDMR, is calibration. Without calibrating sensors, drive mechanisms, communication and algorithms, the most advanced systems can be rendered useless. Of the many types of calibration methods for robotic systems, one specific method is focused on in the second portion of this work: the "hand-eye" calibration problem for robotics. Of specific interest was the AX=XB formulation of the calibration problem in which, the data from sensors on multiple locations of a robotic system can be configured to known reference points such that they can provide correlated data. In this work, two new solution methods for the $AX=XB$ sensor calibration problem are presented, namely, the batch methods and the gradient descent method. The first of these methods combines pose data as a probabilistic "batches" and solves for the missing calibration parameters using probability theory on Lie groups. This solution method is unique, in that, it is able to find the calibration parameters of the system without knowledge of the direct correspondences between readings from separate sensors, knowledge that is required in other solution methods. In the second method, a cost function for the calibration problem is formulated which maps the sensor readings (elements of a Lie group) to a scalar quantity. The calibration parameters are evaluated by following a gradient decent on this cost function. With this method, the evaluated calibration parameters could be constantly updated in real time, continually searching for a new minimum with new incoming data. Ultrasound (US) calibration was chosen to test these solution methods with real data. This data was collected by imaging a known calibration phantom with an US transducer in a series of poses. The two algorithms were applied to this data and it was shown that the new methods were successful, giving error that was well within the margin of other systems. While application to ultrasound calibration is used as a primary example, the solution methods presented can be used for many other fields where the $AX=XB$ sensor calibration problem must be solved. The final methods that were developed use invariant quantities inherent in the structure of the geometric sensor pose data to correct for unknown and missing correspondences in the sensor data streams. These invariants are used in two algorithms, the first of which can realign uniformly asynchronous data and, in some cases, data with gaps. The second algorithm solves for the calibration parameters for most instances of shifts and gaps in the data streams.
Modular Robotics, Team Repair, Ultrasound, Euclidean Geometry, Design