CS 201: Bootstrapping Vehicles: a Formal Approach to Unsupervised Sensorimotor Learning Bases on Invariances, ANDREA CENSI, Caltech

Bootstrapping Vehicles: a Formal Approach to Unsupervised Sensorimotor Learning Bases on Invariances

Andrea Censi

Caltech

Abstract:

Suppose that an agent wakes up in an unknown robotic body, connected to a stream of uninterpreted observations and commands, having zero prior information on its sensors, its actuators, and the external world. How can the agent "bootstrap" a model of its body from scratch, in an unsupervised manner, and use it to perform useful tasks?  This problem sits at the intersection of numerous scientific questions and engineering problems, and can be investigated from various viewpoints.

I am interested in understanding whether this bootstrapping problem can be formalized to the point where it can be solved with the rigour of control theory. For example, currently many learning methods implicitly assume to be able to deal with "generic black boxes", yet it is clear that some assumptions about the semantics of the data are needed---no agent can work if the streams are encrypted. What is missing is a language that allows to formalize the agent's assumptions about the data.  I will show that some classes of semantic assumptions can be considered as the dual of certain group actions that act as fixed "representation nuisances" on the data streams while preserving the controllability/observability properties of the system. The behavior generated by an optimal agent must be invariant to the representation nuisances; thus by characterizing its invariance properties we can describe an agent's semantic assumptions. Moreover, formalizing bootstrapping as rejection of group nuisances integrates naturally with a modular analysis of the various preprocessing layers of a bootstrapping agent.

These ideas will be instantiated at a more concrete level in a "Braitenberg Vehicles" scenario; i.e., considering exteroceptive sensors on a mobile robot (this is approximately half of robotics, the other half being articulated bodies, touch, and manipulation). I will discuss what classes of models can capture the dynamics of three "canonical" robotic sensors (camera, range-finder, field sampler) and the trade-offs of simplicity, prior information needed, and omputational efficiency.

Bio: 

Andrea Censi is a graduate student in Control & Dynamical Systems at the California Institute of Technology. He obtained a M.Eng. degree in Control Engineering and Robotics from La Sapienza University in Rome, Italy, in 2007. He is broadly interested in perception and decision making problems for embodied agents, and in particular in estimation, filtering, and learning in robotics.

Hosted by Prof Stefano Soatto

REFRESHMENTS at 3:45 pm, SPEAKER at 4:15 pm                            

TUESDAY, FEBRUARY 28, 2012 

PLACE: 3400 Boelter Hall

TIME: 4:15 – 5:45 PM