The ability to autonomously plan a course of action to reach a desired goal in the presence of uncertainty is critical for the success of autonomous robotic systems. Autonomous planning typically involves selecting actions that maximize the mission objectives given the available information, such as models of the agent dynamics, environment, available resources, and mission constraints. However, such models are typically only approximate, and can rapidly become obsolete, thereby degrading the planner performance. Classical approaches to address this problem typically assume that the environment has a certain structure that can be captured by a parametric model that can be updated online. However, finding the right parameterization a priori for complex and uncertain domains is challenging because substantial domain knowledge is required. An alternative approach is to let the data provide the insight on the parameterization. This approach leads to Bayesian Nonparametric models (BNPMs), which is a powerful framework for reasoning about objects and relations in settings in which these objects and relations are not predefined. This feature is particularly attractive for missions, such as long-term persistent sensing, for which it is virtually impossible to specify the size of the model and the number of parameters a priori. In such scenarios, BNPMs are especially well suited for integrating data from multiple sensors , and choosing the appropriate model size based on the observed data.

Gaussian processes (GPs) are an example of a widely used BNPM for regression and clustering problems. However, GPs are only one class of a rapidly growing number of BNPMs that capture, for example, hybrid system dynamics with an unknown number for modes or possibly shared features. While GPs have been used with some success in planning and decision-making applications, the use of other types of BNPMs for these scenarios is much less widespread. This is unfortunate because alternate BNPM techniques, such as Dirichlet Processes (DP), Beta Processes (BP), and their hierarchical variants, are potentially much better suited for discrete classification, clustering, and high level planning that involve discrete decisions.

Thus the purpose of this workshop is to provide a mechanism for the statistical machine learning and autonomous decision-making communities to discuss recent results in BNPM and inference techniques with the goal of investigating how the BNPMs can be used to improve the reliability and performance of autonomous planning and decision making systems in a data-rich world. This will be accomplished by bringing together leading experts and practicing engineers in both fields. The workshop will present a mix of invited sessions, contributed talks, poster session by leading experts and active researchers in BNPMs, robotics, and decision making and planning. Furthermore, the workshop schedule is designed to allow for plenty of mingle and question time that is expected to be conducive to merging of ideas and initiating collaborations.