Task Planning (also known as Symbolic Planning or AI Planning) has proved to be a very useful technique to tackle the decision-making problem in robotics. Given a set of task goals, the planner can come up with a set of actions that will reach those goals once executed by the robot.
However, plans are often short-lived when robots execute them, given that the real world is complex, and some actions may fail. A traditional approach is to recompute plans once they fail (replanning), however, computing new plans is often costly. This is an issue in robotics and real-time systems, where users wouldn’t want a robot that stops for some minutes to compute a plan every now and then. Instead of replanning, a solution could be to repair the plan. While some approaches exist [1, 2], none has yet exploited the semantics of the task and the actions. As the actions are meant to be applied in the real world, the meaning of the action is important and may be used not only to repair plans and post-process them, but also to explain them to users. Moreover, plans involving certain actions and tasks that are not accompanied by a real-world context cannot be guaranteed to be safe or trustworthy for users.
While a full specification of task and action semantics is cumbersome due to the size and complexity of open domains, some ongoing efforts are addressing the also general, but more manageable domain of common sense. For example, OpenCyc has been running for decades to “assemble a comprehensive ontology and knowledge base that spans the basic concepts and rules about how the world works” . More recently, the knowledge graph community has advanced ground in integrating various knowledge bases (e.g. ATOMIC, ConceptNet, FrameNet, Roget, Visual Genome, Wikidata, WordNet) of common-sense knowledge, in a hyper-relational graph called Common Sense Knowledge Graph  (CSKG). A large number of the symbolic representations (e.g. concepts, relations, rules, etc.) in CSKG are relevant for, and could be used as semantic descriptions of, tasks and actions in robot planning.
In this project, we propose to combine the ideas of Symbolic Planning for decision-making in robotics with explicit representations of common-sense knowledge in knowledge graphs for safer planning. The idea is that such a combination can leverage contextual descriptions of domains and use common-sense reasoning to avoid plans containing actions or tasks with the potential of being unsafe or untrustworthy. Furthermore, this may also allow to not only improve plans that might not be executable due to semantic constraints, but also to change them in a way that enhances user trust in the robotic system. The symbolic nature of common-sense knowledge graphs such as the CSKG can provide a layer of explainability ensuring that plans can be understood and debugged by humans, creating feedback loops between the planner and the knowledge graph.
More specifically, this project:
- Assesses common-sense knowledge in explicit symbolic representations, such as those provided by CSKG and other related datasets, as reliable sources of semantic information for robot planning.
- Develops new planning algorithms that leverage common-sense knowledge graphs and common-sense reasoning to propose semantically rich, explainable plans for robots.
- Evaluates the performance, safety and trustworthiness of these implementations by comparing them with existing approaches that do not exploit common sense.