Seamless Levels of Abstraction for Robot Cognition

The usual approach to implement a robotic platform capable of executing human-like tasks is to combine methods from artificial intelligence with robotic mechanisms using different representations and search techniques. This makes integration and consistency checking between techniques very complicated, preventing the accurate predictions of the consequences of actions in real scenarios and, hence, the generation of successful plans. Our main claim is that, in order to reliably and accurately predict the consequences of actions, the symbolic descriptions needed by artificial intelligence methods for making decisions should be evaluated directly from the continuous variables of the physical phenomena they represent. We propose a unified representation that permits this evaluation. We associate to predicates and planning operators two probability density models representing the physical instances where the evaluations are positive or negative. The density models associated to predicates permit an accurate estimation of the probability that the predicate takes value true or false for a given physical configuration. This permits improved descriptions of the initial states for planning, which is fundamental for the generation of successful plans. The density models associated to planning operators, in turn, permit an accurate estimation of how probable the operator will be successfully executed in a given physical configuration of the current situation, which is fundamental for the deliberation process to generate feasible plans. As such a system cannot be complete hand-designed and, in particular, to adapt to unpredicted situations typical of real, human-like scenarios, we define an incremental learning method that generates planning operators and predicates while the task is executed under the guidance of a human teacher. Using the unified representation, we define a probabilistic planning approach that uses these enhanced prediction capabilities to generate reliable plans for the execution of human-like tasks in real robotic scenarios. We will evaluate our approach against the state of the art approaches using appropriate evaluation measures.

Everyday human-like scenarios are highly unstructured and unpredictable. For a robot to operate autonomously in such scenarios, the traditional approach is to combine tools from two different fields: Artificial intelligence (AI) and robotics. AI task planning approaches are used to automatically generate sequences of abstract instructions, called task plans, which define the steps to complete a task from a given initial situation. Robotic motion planning approaches, in turn, are used to transform each abstract instruction into robot motions for task execution. However, AI and robotic techniques are historically incompatible. They were conceived independently for different purposes using different representations and search techniques. The abstract descriptions of AI planning methods make it difficult to consistently represent real-valued physical constraints that must be taken into account to successfully execute a task. This lack of compatibility makes combining AI and robotics techniques in a single robotic architecture a great challenge, as reflected in the limited results obtained to date. This project addresses this problem by proposing a unified representation of physical constraints in terms of object positions and orientations, in an object-centric approach compatible with both AI task and robotic motion planning. At the task planning level, object parameters are used to define physical constraints in terms of functional parts of objects. For example, to pour water into a glass, the top of the glass must be open and pointing upwards, the left side of the bottle must be clear to be grasped by the robot's hand, and so on. This allows AI planning to produce task plans that satisfy the physical constraints for their execution. At the motion planning level, the physical constraints associated with functional parts are transformed back into positions and orientations of objects for the generation of robot motions, such as the required position and orientation of the hand to grasp a bottle to pour afterwards. The proposed object-centric approach is complemented with learning mechanisms that automatically generate from human demonstrations those constraints that cannot be encoded in terms of functional object parts, such as constraints to avoid collisions while the robot is moving. Our contributions leverage the efficiency of AI planning methods to generate physically feasible task plans that can be directly transformed into robotic movements with a high success rate and without intensive computation. These are fundamental requirements for robots to operate in everyday scenarios, where simple robotic tasks such as picking a medicine off a shelf or pouring a glass of water can significantly improve the daily life of an elderly person living alone or a person with limited mobility.