Editorial: Robots that Learn and Reason: Towards Learning Logic Rules from Noisy Data
Plínio Moreno, Alexandre J. M. Bernardino, José Santos-Victor, Rodrigo Ventura, Kristian Kersting · Frontiers in Robotics and AI · 2021
Robots that Learn and Reason: Towards Learning Logic Rules from Noisy DataFrom the early developments of AI applied to robotics by Hart et al. (1968), Duda and Hart (1972) and Lozano-Pérez and Wesley (1979), higher level commands were grounded to real world sensing by carefully design algorithms, which provide a link between the abstract predicates and the sensors and actuators.In order to have fully autonomous robots that learn by exploration and by imitation, the grounding algorithms between the higher-level predicates and the lower-level sensors and actuators should be discovered by the robot.Previous and recent efforts on robotics aim to discover and/or learn these intermediate layer commands, which must cope with discrete and continuous data.The main objective of this Research Topic is to advance on learning logic rules from noisy data.We have four articles that address: Logic rules that cope with states that are not directly observable by the sensing modalities; learning rules that represent object properties and their functionalities, which are grounded to the particular robot experience; learning low-level robot control actions that fulfill a set of abstract predicates in a two-level planning approach; learning to develop skills in a robotic playing scenario by composing a set of behaviors.In the following, we introduce the four articles and their contributions to rule learning in presence of noisy data.The article by Zuidberg Dos Martires et al. introduces a general-purpose semantic object tracker, which anchors sensory input to object representations in a probabilistic manner.The semantic tracker is able to estimate the object position in the presence of total occlusions of the object by using a multi-modal probability distribution, which is modeled as a distributional logic tree.In addition, the tree structure of the multi-modal distribution is estimated from observations, which provides a set of data-driven clauses that can be augmented with recursive ones.This methodology allows to model objects that can be occluded by objects that are themselves occluded.The article by Thosar et al. defines measurable object properties and their functionalities, focusing on the robot-centered estimation of the object properties for multiple categories.Regarding the object properties, the robot executes actions that interact with the objects, measuring the effects on the objects through proprioceptive changes in the robot and external sensors.The robot-centric conceptual knowledge is obtained from unsupervised clustering of the numerical values, where each cluster represents a qualitative property (i.e., concept).Regarding the object functionalities, these are selected according to a set of predefined tasks, where a quantitative value of the functionality is obtained from vision-based sensors.The object properties and their functionalities are designed for selecting a tool substitute, which is addressed as a classification problem.The learning algorithm selects tool substitutes very similar to human experts, showing the validity of the approach.This methodology allows to select tool substitutes in a robotcentric manner, based on unsupervised discovery of qualitative object properties.