My project while at the AI Lab was called 'A robot thinking ahead' and it was part of 'From locomotion to cognition', a Swiss National Science Foundation project (Grant Nr. 200020-122279/1). The objective of my project was to investigate the use of forward models in a quadruped robot.
In my project I developed a bio-inspired control architecture that allows a mobile robot to: (1) learn a model of its own action repertoire (a forward model); (2) learn a model of an object (prey) it is seeking; (3) combine the forward model and the prey model to seek the prey.
This work has been published in the following paper:
We present a bio-inspired control method for moving-target seeking with a mobile robot, which resembles a predator-prey scenario. The motor repertoire of a simulated Khepera robot was restricted to a discrete number of 'gaits'. After an exploration phase, the robot auto- matically synthesizes a model of its motor repertoire, acquiring a forward model. Two additional components were introduced for the task of catching a prey robot. First, an inverse model to the forward model, which is used to determine the action (gait) needed to reach a desired location. Second, while hunting the prey, a model of the prey's behavior is learned online by the hunter robot. All the models are learned ab initio, without assumptions, work in egocentric coordinates, and are probabilistic in nature. Our architecture can be applied to robots with any physical constraints (or embodiment), such as legged robots.
The following video shows the hunter robot (black) trying to catch the prey robot (red) using only the reactive model (i.e. the inverse model to the forward model).
The following video shows the hunter robot trying to catch the prey robot using (as well as the inverse model) a model of the prey's pose transitions in order to predict future poses.
The following video shows the same but now the robots are in an open environment and these models are no longer enough to catch the prey.
The following video shows that the hunter is able to catch the prey again even in an open environment by using a planning model.