About this Research Topic
This Research Topic is based on outputs from the workshop 'Robots that Learn and Reason' at the International Conference on Intelligent Robots (IROS) 2018. However, we would also welcome spontaneous submissions not associated with this workshop, which should fit the scope of the Research Topic.
The main objective of this Research Topic is to provide new insights into the association of rules (i.e. concepts, abstractions) to sensorimotor robots' data, in the context of the recent developments presented in the articles “Learning explanatory rules from noisy data“, and “Learning abstract hierarchical compositional visual concepts“. These recent developments have shown that is possible, on the one hand, to learn (human-readable) explicit symbolic rules while being robust to noisy and ambiguous data; and on the other, learn associations between visual input and logical recombination operators in a hierarchical, fully grounded and with very little supervision. The connections between logic programming and robotics have been studied for a while, but the association between noisy and ambiguous data and symbolic rules has been performed by humans, using educated guesses and previous experiences. This is similar to the previous approaches in computer vision, where the image features were usually developed and selected by humans. Currently, deep learning approaches work in an end-to-end fashion so the image feature selection is done by the learning algorithm as well. Thus, the big question to be discussed is: “Are robots ready to learn symbolic rules and their association to noisy and ambiguous data in an end-to-end manner?”.
Addressing this question has several issues to be considered, so we propose to have two sides:
(i) The pragmatic and engineering approach, and
(ii) the long-term quest of a solution.
From the point of view of the pragmatic and engineering approach, it is very important to understand: (i) what are the symbolic rules and what kind of symbolic rules can be learned from sensors such as lasers, sonars, tactile sensors, etc?; and (ii) which type of robotics tasks could benefit from such logic programming approach? (i.e navigation, manipulation, interaction with humans). From a long-term quest of a solution, it is crucial to discuss what theoretical and technological insights are missing in order to reach end-to-end associations of symbolic rules to noisy and ambiguous data. Furthermore, in a more speculative discussion, the invited speakers and the audience will comment on: Will the robots be able to discover the symbolic rules from noisy and ambiguous data?
We expect submission that address one or more the above questions, encouraging roboticist and AI enthusiasts to present their current work on Inductive Logic Programming, Probabilistic Programming Languages and Relational learning to robotics problems. We especially would like to have papers on rule learning and rule discovery for robotics applications, but we welcome works in other areas such as rule-based planning.
Articles should include address at least one of the following topics:
• Rule learning for autonomous robots
• Rule learning from noisy data
• Rule discovery in robotic domains
• Visual grounding of concepts
• Sensorimotor grounding of concepts
• Probabilistic programming languages
• Inductive Logic programming
The main objective of this Research Topic is to provide new insights into the association of rules (i.e. concepts, abstractions) to sensorimotor robots' data, in the context of the recent developments presented in the articles “Learning explanatory rules from noisy data“, and “Learning abstract hierarchical compositional visual concepts“. These recent developments have shown that is possible, on the one hand, to learn (human-readable) explicit symbolic rules while being robust to noisy and ambiguous data; and on the other, learn associations between visual input and logical recombination operators in a hierarchical, fully grounded and with very little supervision. The connections between logic programming and robotics have been studied for a while, but the association between noisy and ambiguous data and symbolic rules has been performed by humans, using educated guesses and previous experiences. This is similar to the previous approaches in computer vision, where the image features were usually developed and selected by humans. Currently, deep learning approaches work in an end-to-end fashion so the image feature selection is done by the learning algorithm as well. Thus, the big question to be discussed is: “Are robots ready to learn symbolic rules and their association to noisy and ambiguous data in an end-to-end manner?”.
Addressing this question has several issues to be considered, so we propose to have two sides:
(i) The pragmatic and engineering approach, and
(ii) the long-term quest of a solution.
From the point of view of the pragmatic and engineering approach, it is very important to understand: (i) what are the symbolic rules and what kind of symbolic rules can be learned from sensors such as lasers, sonars, tactile sensors, etc?; and (ii) which type of robotics tasks could benefit from such logic programming approach? (i.e navigation, manipulation, interaction with humans). From a long-term quest of a solution, it is crucial to discuss what theoretical and technological insights are missing in order to reach end-to-end associations of symbolic rules to noisy and ambiguous data. Furthermore, in a more speculative discussion, the invited speakers and the audience will comment on: Will the robots be able to discover the symbolic rules from noisy and ambiguous data?
We expect submission that address one or more the above questions, encouraging roboticist and AI enthusiasts to present their current work on Inductive Logic Programming, Probabilistic Programming Languages and Relational learning to robotics problems. We especially would like to have papers on rule learning and rule discovery for robotics applications, but we welcome works in other areas such as rule-based planning.
Articles should include address at least one of the following topics:
• Rule learning for autonomous robots
• Rule learning from noisy data
• Rule discovery in robotic domains
• Visual grounding of concepts
• Sensorimotor grounding of concepts
• Probabilistic programming languages
• Inductive Logic programming
Keywords: Rule Learning, Rule Discovery, Concept Discovery, Sensorimotor Concept Grounding
Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
