Language Acquisition and Robotics Group
at the University of Illinois at Urbana-Champaign
Humans are able to acquire complex, goal-directed movements from observation of other humans. This process begins with simple inborn reflexes which help bootstrap learning, such as grasping. This progresses to the use of motor babbling to learn basic muscle control, and then to the acquisition of motor primitives. The library of motor primitives amassed by early childhood allows the fast learning of a wide range of actions. Our goal is to replicate this process in the iCub. Some of the algorithmic approaches being investigated are hierarchies of self-organizing maps and learning dynamics of topographically arranged neural networks. Motor imitation will allow robots to adaptively learn skills and social cues from interaction with human instructors and provide insight in to possible mechanisms for human motor control.
For creation of an artificial agent that is capable of using language naturally, models which only manipulate symbols or classify speech are ineffective. The semantic information which language conveys must be grounded in the agent’s complete sensorimotor experience. Before the process of language learning can even begin, it is first necessary convert the broad array of information which we experience into a compact representation which is accesible to the many different functions of cognition. For robots, this means developing computational structures capable of encoding sensory information.
Dynamical Models of Associative Memory
The lab is also developing ab-initio nonlinear dynamical multi-scale model of associative memory. The model will enable a robust way of encoding noisy, complex, multi- dimensional, multi-modal, real-time input streams from the external world (speech, visual scenes, joint angles, etc.) into a low-dimensional trajectory form.
Scale 0 consists of the canonical, conductance-based spiking Hodgkin- Huxley (HH) neuron model.
Scale 1 consists of components that are made up of large populations of HH neurons whose topological structure evolves according to a Hebbian- plasticity rule based on synchronous firing. The component’s state is captured by the variance of phase synchrony for the population. Many such components are sparsely connected to form a large network, whose state can be captured by the n-tuple consisting of the individual states of each member component.
Scale 2 takes the state of the overall network and upon examining the particular interrelationships of each component (determining how the state of one component affects the state of others) is able to generate a class of trajectories that is multistationary and stable periodic. Such a class we consider a memory; the encoding of many such memories leads to the creation of a robust associative memory.
Fine Motor ControlMotor function is a necessary component of our ability to interact with the world, as well to understand it. Before beginning to approach the problems of how humans are able to execute precision motor tasks on the cognitive level, it is helpful to understand such tasks at a basic physical level as well. Using traditional engineering methods for control and mechanical modeling, part of our research involves testing the physical abilities of our robotic platform, and how these compare to our own abilities for tasks such as walking, reaching, grasping, and placing objects.