Search

Brain-computer interfaces (BCIs)—systems that can record neural activity and translate them into commands for computer systems—are sufficiently advanced to allow users to volitionally guide them through simple tasks. Contemporary BCI research focuses on squeezing additional functionality out of standardized paradigms, be it achieving more bits per second, increased degrees of freedom, or increasing accuracy. While these studies have shown marginal advancements in recent years, our lack of understanding concerning the underlying neurophysiology continues to be the limiting factor in BCI development. In this chapter, we propose turning the way research is done on BCI systems on its head; instead of using our understanding of neural signals to incrementally advance the state of brain-machine interfaces, we apply a BCI system as a form of experimental control to study changes in neural activity. By using current BCI systems as a tool for neuroscientific study, we can probe the underlying neuroanatomy in novel, behaviorally controlled ways.

The performance of non-invasive electroencephalogram-based (EEG) brain-computer interfacing (BCI) has improved significantly in recent years.However, remaining challenges include the non-stationarity and the low signal-to-noise ratio (SNR) of the EEG, which limit the bandwidth and hence the available applications. In this paper, we review ongoing research in our labs and introduce novel concepts and applications. First, we present an enhancement of the 3-class self-paced Graz-BCI that allows interaction with the massive multiplayer online role playing game World of Warcraft. Second, we report on the long-term stability and robustness of detection of oscillatory components modulated by distinct mental tasks. Third, we describe a scalable, adaptive learning framework, which allows users to teach the BCI new skills on-the-fly. Using this hierarchical BCI, we successfully train and control a humanoid robot in a virtual home environment.

Programming a humanoid robot to perform an action that takes into account the robot’s complex dynamics is a challenging problem. Traditional approaches typically require highly accurate prior knowledge of the robot’s dynamics and the environment in order to devise complex control algorithms for generating a stable dynamic motion. Training using human motion capture (mocap) data is an intuitive and flexible approach to programming a robot but direct usage of kinematic data from mocap usually results in dynamically unstable motion. Furthermore, optimization using mocap data in the high-dimensional full-body joint-space of a humanoid is typically intractable. In this chapter, we purposes a new model-free approach to tractable imitation-based learning in humanoids by using eigenposes.

A prerequisite for achieving brain-like intelligence is the ability to rapidly learn new behaviors and actions. A fundamental mechanism for rapid learning in humans is imitation: children routinely learn new skills (e.g., opening a door or tying a shoe lace) by imitating their parents; adults continue to learn by imitating skilled instructors (e.g., in tennis). In this chapter, we propose a probabilistic framework for imitation learning in robots that is inspired by how humans learn from imitation and exploration. Rather than relying on complex (and often brittle) physics-based models, the robot learns a dynamic Bayesian network that captures its dynamics directly in terms of sensor measurements and actions during an imitation-guided exploration phase. After learning, actions are selected based on probabilistic inference in the learned Bayesian network. We present results demonstrating that a 25-degree-of-freedom humanoid robot can learn dynamically stable, full-body imitative motions simply by observing a human demonstrator.

Imitation-based learning is a general mechanism for rapid acquisition of new behaviors in autonomous agents and robots. In this paper, we propose a new approach to learning by imitation based on parameter learning in probabilistic graphical models. Graphical models are used not only to model an agent's own dynamics but also the dynamics of an observed teacher. Parameter tying between the agent-teacher models ensures consistency and facilitates learning. Given only observations of the teacher's states, we use the expectation-maximization (EM) algorithm to learn both dynamics and policies within graphical models. We present results demonstrating that EM-based imitation learning outperforms pure exploration-based learning on a benchmark problem (the FlagWorld domain). We additionally show that the graphical model representation can be leveraged to incorporate domain knowledge (e.g., state space factoring) to achieve significant speed-up in learning.

We argue that human vision has natural timescales, and that models of human vision at these different timescales are qualitatively different. In particular, at the timescale of a few seconds, human vision can be modeled in terms of two primitive functional routines. A “what” routine determines object identity from a segmented input and a “Where” routine determines the retinal location of a desired object. More complicated functions can be composed from these two. In particular, a complicated visuo-motor task such as copying can be described in terms of these two routines. The primary subroutine needed is one that computes the relationship of the parts of an object with respect to an object-centered frame.

The location of objects in images is difficult owing to the view variance of geometric features but can be determined by developing viewinsensitive descriptions of the intensities local to image points. Viewinsensitive descriptions are achieved in this work by describing points in terms of the responses of steerable filters at multiple scales. Owing to the use of multiple scales, the vector for each point is, for all practical purposes, unique, and thus can be easily matched to other instances of the point in other images. We show that this method can be extended to handle the case where the area near a point of interest is partially occluded. The method uses a description of the occluder in the form of a template that can be obtained easily via active vision systems using a method such as disparity filtering.