Reasoning and understanding grasp affordances for robot manipulation

Abstract

This doctoral research focuses on developing new methods that enable an artificial agent to grasp and manipulate objects autonomously. More specifically, we are using the concept of affordances to learn and generalise robot grasping and manipulation techniques. [75] defined affordances as the ability of an agent to perform a certain action with an object in a given environment. In robotics, affordances defines the possibility of an agent to perform actions with an object. Therefore, by understanding the relation between actions, objects and the effect of these actions, the agent understands the task at hand, providing the robot with the potential to bridge perception to action. The significance of affordances in robotics has been studied from varied perspectives, such as psychology and cognitive sciences. Many efforts have been made to pragmatically employ the concept of affordances as it provides the potential for an artificial agent to perform tasks autonomously. We start by reviewing and finding common ground amongst different strategies that use affordances for robotic tasks. We build on the identified grounds to provide guidance on including the concept of affordances as a medium to boost autonomy for an artificial agent. To this end, we outline common design choices to build an affordance relation; and their implications on the generalisation capabilities of the agent when facing previously unseen scenarios. Based on our exhaustive review, we conclude that prior research on object affordance detection is effective, however, among others, it has the following technical gaps: (i) the methods are limited to a single object ↔ affordance hypothesis, and (ii) they cannot guarantee task completion or any level of performance for the manipulation task alone nor (iii) in collaboration with other agents. In this research thesis, we propose solutions to these technical challenges. In an incremental fashion, we start by addressing the limited generalisation capabilities of, at the time state-of-the-art methods, by strengthening the perception to action connection through the construction of an Knowledge Base (KB). We then leverage the information encapsulated in the KB to design and implement a reasoning and understanding method based on statistical relational leaner (SRL) that allows us to cope with uncertainty in testing environments, and thus, improve generalisation capabilities in affordance-aware manipulation tasks. The KB in conjunctions with our SRL are the base for our designed solutions that guarantee task completion when the robot is performing a task alone as well as when in collaboration with other agents. We finally expose and discuss a range of interesting avenues that have the potential to thrive the capabilities of a robotic agent through the use of the concept of affordances for manipulation tasks. A summary of the contributions of this thesis can be found at: https://bit.ly/grasp_affordance_reasoning