Logo: University of Southern California

Events Calendar

  • Bootstrapping Vehicles: a Formal Approach to Unsupervised Sensorimotor Learning Based on Invariance

    Thu, Mar 07, 2013 @ 03:30 PM - 05:00 PM

    Computer Science

    Conferences, Lectures, & Seminars


    Speaker: Andrea Censi, Caltech

    Talk Title: CS Colloquium: Andrea Censi (CalTech)

    Series: CS Colloquium

    Abstract: Imagine you are a brain that wakes up in an unknown (robotic) body. You are connected to two streams of uninterpreted observations and commands. You have zero prior information on the body morphology, its sensors, its actuators, and the external world. Would you be able to "bootstrap" a model of your body from scratch, in an unsupervised manner, and use it to perform useful tasks? This bootstrapping problem sits at the intersection of numerous scientific questions and engineering problems. Biology gives us a proof of existence of a solution, given that the neocortex demonstrates similar abilities.

    I am interested in understanding whether the bootstrapping problem can be formalized to the point where it can be solved with the rigour of control theory. I will discuss a tractable subset of the set of all robots called the "Vehicles Universe", which I consider a pimped-up version, with modern sensors, of Braitenberg's Vehicles. I will show that the dynamics of three "canonical" robotic sensors (camera, range-finder, field sampler) are very similar at the "sensel" level. I will present classes of models that can capture the dynamics of those sensors simultaneously and allow exactly the same agent to perform equivalent spatial tasks when embodied in different robots. I will discuss immediate applications to intrinsic sensor calibration and fault detection.

    A key concern of mine is to precisely characterize the "assumptions" of the agents. I will show that assumptions regarding the representation of the data can be described by the largest group of transformations on observations/commands to which the agent behavior is invariant. This suggests that one of the basic concerns of a bootstrapping agent is being able to reject these "representation nuisances".

    Reference: the homonymous dissertation, available at http://purl.org/censi/2012/phd


    Biography: Andrea Censi is a postdoctoral scholar in Computing and Mathematical Sciences at the California Institute of Technology. He received the Laurea and Laurea Specialistica degrees (summa cum laude) in control engineering and robotics from Sapienza University of Rome, Italy, in 2005 and 2007, respectively, and a Ph.D. in Control & Dynamical Systems from the California Institute of Technology in 2012. He is broadly interested in perception and decision making problems for natural and artificial embodied agents, and in particular in estimation, filtering, and learning in robotics.

    Website: http://andrea.caltech.edu/


    Host: Fei Sha

    Location: Seaver Science Library (SSL) - 150

    Audiences: Everyone Is Invited

    Posted By: Assistant to CS chair

    OutlookiCal

Return to Calendar