ActAR - Aktionsbewusstsein für kognitive Roboter
Zusammenfassung der Projektergebnisse
Performing everyday tasks requires robots to continually make decisions about which actions to execute, and also how to best parameterize these actions for given a specific task context. Making such decisions is a complex process, especially as many relevant factors for the task context are often not known initially, but only become observable during task execution. Mobile manipulation is a good example. Consider the fundamental mobile manipulation task of navigating to a position in order to grasp an object from a table. Successfully executing this task requires the robot to decide where to stand in order to pick up the object, which hand(s) to use, how to reach for it, and many others. These decisions depend on the location of the robot, the location of the object, and the uncertainties the robot has about these locations. Through our research, we are showing ways of making the robot behaviour more reliable, flexible and efficient in such setting. We propose the concept of Action-Related Place (ARPLACE) as a powerful and flexible representation of task-related place in the context of mobile manipulation. ARPLACE represents base places not as a single position, but rather as a collection of positions, each with an associated probability that the manipulation action will succeed when standing there. ARPLACES are generated using a predictive model that is acquired through experience-based learning. As this model is learned from observed successful and failed action executions, it is grounded in the robot's actual behavior. When executing the task, the robot instantiates an ARPLACE, and bases its decisions on this ARPLACE, rather than choosing one specific goal position based only on the initial knowledge about the task context. Having the respective place concept instantiation represented explicitly during the course of action enables the robot to reconsider and reevaluate the decision whenever new information about the task context comes in. To show the advantages of this least-commitment approach, we integrate the ARPLACE representation in a symbolic transformational planner. Our empirical evaluation demonstrates that using ARPLACE leads to more robust and efficient mobile manipulation in the face of state estimation uncertainty. Apart from the measures that directly relate to the task performance itself such as the mentioned reliability, flexibility and efficiency, we want our robots to be able to explain what they were doing, how they did it and why. Understanding the motives behind actions is not only helpful for introspection by the human developer but rather provides the mechanisms for complex monitoring and failure handling that goes beyond local failures and exceptions. Using semantic annotations of plans, we can provide an expressive formulation of failures and error patterns in first order logic. Reasoning on top of a fast logging mechanisms and in-time calculation of predicates provides awareness of the agent's actions. In the following we point out the key findings and main contributions of the conducted research. There are three main areas of contribution, the concept of ARPlaces itself. Action-awareness through Reasoning about Logged Plan Execution Traces and the use of both techniques in a Transformational Planner. In the context of ARPlaces, we have developed a system that enables robots to learn actionrelated places from observed experience (either real or simulated), and reason with these places to generate robust, flexible, least-commitment plans for mobile manipulation. ARPLACE is modeled as a probability distribution that maps locations to the predicted outcome of an action. ARPLACE are a flexible representation of place for least-commitment decision making in mobile manipulation. Furthermore, the learned model is very compact, with only 2 (deformation) parameters, which are directly related to task-relevant parameters. Querying the model on-line is therefore very efficient. This is an advantage of compiling experience into compact models, rather than running a novel search for each situation. With respect to Reasoning about Logged Plan Execution Traces, we presented an extension to CRAM that implements high performance and accurate reasoning mechanisms. These allow to answer complex questions about plan execution, the intention of the robot, the reason for failures and the belief state of the robot. This is achieved by creating an extensive execution trace and mechanisms to query it through a first-order representation. Since this system can not only be used offline but also during plan execution, i.e. within control routines, it enables deep and complex failure handling mechanisms based on descriptions of failures in the first-order representation. Both ARPlaces and Reasoning on Execution Traces were used in the Transformational Planner within the CRAM environment. To that end Execution Traces are queried to identify a performance flaws in a past execution and fix the default plan for similar occasions in the future. Generic Patterns within plans that can be identified in the execution traces, such as grasping two items in sequence. Once a pattern is found and the performance in this context is assessed and Transformation Rules are checked for their applicability. The ARPlaces are used in this case to predict the outcome of the plan after Transformation by projecting the plan outcome on-line. If a significant performance gain is expected, the default plan for similar occasions is changed to the transformed plan. This loop is closed due to the ARPlaces' ability to consider learned real data, enabling to automatic correction of actual performance losses by adaptation of the prediction model to the real performance of the system, allowing for the decisions made to be more informed over time. Pick-and-place tasks are very common across many applications scenarios. Therefore, the developed methods can be employed within a wide range of daily tasks. In general the system could be used for all kinds of fetch-and-delivery tasks, e.g., within a household, a hospital, and a factory setting. At the same time, the developed mechanisms are not limited to pick-and-place tasks at all. We ourselves are planning to extend the approach in several directions. We are in the process of including ARPLACE in a more general utility-based framework, in which the probability of success is only one of the aspects of the task that needs to be optimized. New utilities, such as execution duration or power consumption, are easily included in this framework, which enables the robot to trade off efficiency and robustness on-line during task execution. We are also going to apply our approach to more complex scenarios and different domains. For instance, we are learning higher-dimensional ARPLACE concepts, which take more aspects of the scenario into account, i.e. different object sizes and objects that require different types of grasps. Instead of mapping specific objects to places, we will map object and grasp properties to deformation modes. We are also investigating extensions and other machine learning algorithms that will enable our methods to generalize over this larger space. Objects which require very different grasps, such as using two hands to manipulate them, will require more sophisticated methods for acquiring and reasoning about place.
