|Pierre E. Dupont
Autonomous task performance is the ultimate goal of robotics. To this end, enhancing robot perception, the ability of a robot to recognize and model its environment, is essential. This thesis presents design tools for contact-based perceptual systems applicable to manipulation tasks which can be described by sequences of contact states between rigid objects. The fundamental component of such systems is a contact state estimator. This estimator uses sensor data collected as objects are manipulated to determine the sequence of actual contact states from a network of possible contact states. In the process, it also estimates the parameters, e.g., geometric parameters, of the individual contact state models.
In this thesis, a rigorous approach to contact state estimator design is proposed which involves characterizing four properties of a given contact state network, set of sensors, and associated contact state models. These properties are: (1) the distinguishability of contact states, (2) the observability of each contact state inside the state network, (3) the identifiability of each contact state model’s parameters, and (4) the excitability of the sensor inputs to permit estimation of all the parameters associated with a contact state model. The first two properties address the feasibility of contact state estimation, while the last two address the feasibility of estimating the parameters of the individual contact states.
The major contribution of this thesis is the development of a unified analytic approach to testing the distinguishability of contact states and the identifiability of their parameters. The testing method is applicable to any contact state model regardless of the chosen sensing modality. The concept of contact observability is also introduced as a forward projection of the parameter history associated with the execution of the task. The effect of the sensor signals on the parameters is analyzed by studying the invertibility of an excitability matrix representing the relationship between the structure of the contact states and the sensor signals. Finally, an implementation of a contact state estimator is presented using a hidden Markov model to combine a nonlinear least squares estimator with prior information from a contact state network.