Evan Drumwright

Exemplar-based primitives for humanoid movement classification and control

We present a unified methodology for humanoid robot control and activity, classification using motor primitives (Mataric, M, 2002), computationally efficient behaviors capable of perception and control. These primitives constitute a vocabulary for humanoid control capable of generating a large variety of complex movement through sequencing and superposition. We demonstrate how such primitives can be automatically derived from human motion-capture data, how they can be used to construct upperbody controllers, and how they can be applied to classification of observed humanoid behavior in real time.

A Fast and Stable Penalty Method for Rigid Body Simulation

Two methods have been used extensively to model resting contact for rigid-body simulation. The first approach, the penalty method, applies virtual springs to surfaces in contact to minimize interpenetration. This method, as typically implemented, results in oscillatory behavior and considerable penetration. The second approach, based on formulating resting contact as a linear complementarity problem, determines the resting contact forces analytically to prevent interpenetration. The analytical method exhibits an expected-case polynomial complexity in the number of contact points and may fail to find a solution in polynomial time when friction is modeled. We present a fast penalty method that minimizes oscillatory behavior and leads to little penetration during resting contact; our method compares favorably to the analytical method with regard to these two measures while exhibiting much faster performance both asymptotically and empirically.

Toward Interactive Reaching in Static Environments for Humanoid Robots

Reaching is a critical task for humanoid robots, requiring the application of state-of-the-art algorithms for motion planning and inverse kinematics. Practical algorithms for solving these subproblems are currently not complete, offering resolution completeness or probabilistic completeness instead. While this lesser provision of completeness is acceptable in many cases, naively combining state-of-the-art approaches in motion planning and inverse kinematics can lead to a method that provides no measure of completeness. We present a probabilistically complete solution to the reaching problem for humanoid robots in static environments, and evaluate it against two other methods using a kinematically simulated humanoid in a virtual environment

Virtual test tubes: a new methodology for computing

Biomolecular computing (BMC) aims to capture the innumerable advantages that biological molecules have gained in the course of millions of years of evolution to perform computation unfeasible on conventional electronic computers. While biomolecules have resolved fundamental problems as a parallel computer system that we are just beginning to decipher, BMC still suffers from our inability to harness these properties to bring biomolecular computations to levels of reliability, efficiency and scalability that are now taken for granted with conventional solid-state based computers. The authors explore an alternative approach to exploiting these properties by building virtual test tubes in software that would capture the fundamental advantages of biomolecules, in the same way that evolutionary algorithms capture in silico the key properties of Darwinian evolution. We use a previously built tool, Edna, to explore the capabilities of the new paradigm.

Neurofuzzy recognition and generation of facial features in talking heads

We show that fuzzy neural nets can rate the emotional feedback message implicit in human facial expressions at a level comparable to human performance. We also identify primitive and derived features in facial expressions and animations. We also report on a partial solution to the difficult inverse problem of generating naturalistic facial expressions from summary abstract descriptions of the emotional feedback to be conveyed.

Expanding task functionality in established humanoid robots

Many humanoid robots like ASIMO are built to potentially perform more than one type of task. However, the need to maintain a consistent physical appearance of the robot restricts the installation of additional sensors or appendages that would alter its visual identity. Limited battery power for free-moving locomotive robots places temporal and spacial complexity limits on the algorithms we can deploy on the robot. With these conditions in mind, we have developed a distributed robot architecture that combines onboard functionality with external system modules to perform tasks involving interaction with the environment. An information model called the Cognitive Map organizes output produced by multiple perceptual modules and presents a common abstraction interface for other modules to access the information. For the planning and generation of motion on the robot, the Task Matrix embodies a task abstraction model that maps a high level task description into its primitive motions realizable on the robot. Our architecture supports different control paradigms and information models that can be tailored for specific tasks. We demonstrate environmental tasks we implemented with our system, such as pointing at moving objects and pushing an object around a table in simulation and on the actual ASIMO robot.

A robust and tractable contact model for dynamic robotic simulation

Existing contact modeling in rigid body simulation is inadequate for robotics: no algorithms guarantee both convergence and nonpenetration at multiple contact points in the presence of Coulomb friction. We present a convex optimization based algorithm that models simultaneous contact at multiple points, ensures nonpenetration, and yields Coulomb friction effects. An example of simulated robotic grasping shows that the proposed algorithm is robust where most other methods fail.

The task matrix: an extensible framework for creating versatile humanoid robots

The successful acquisition and organization of a large number of skills for humanoid robots can be facilitated with a collection of performable tasks organized in a task matrix. Tasks in the matrix can utilize particular preconditions and in conditions to enable execution, motion trajectories to specify desired movement, and references to other tasks to perform subtasks. Interaction between the matrix and external modules such as goal planners is achieved via a high-level interface that categorizes a task using its semantics and execution parameters, allowing queries on the matrix to be performed using different selection criteria. Performable tasks are stored in an XML-based file format that can be readily edited and processed by other applications. In its current implementation, the matrix is populated with sets of primitive tasks (e.g. reaching, grasping, arm-waving) and macro tasks that reference multiple primitive tasks (pick-and-place and facing-and-waving)

Training a neurocontrol for talking heads

Talking heads are anthropomorphic representations of a software agent used to facilitate interaction with human users. Talking heads have been commonly programmed and controlled by ontologies designed according to intuitive and heuristic considerations that may have little to do with the applications at hand, and are thus probably not truly expressive, meaningful or ergonomic to human users. Here we present preliminary results in the design and training of an autonomous neural control that is capable of generating facial expressions that convey meaningful emotional content to users in the context of tutoring sessions on a particular domain (computer literacy) on a continuous scale of negative, neutral, and positive feedback. The ultimate goal of the project is to have the control autonomously synchronize the movements of facial features in lips, eyes and eyebrows in order to produce facial animations that are not only valid and meaningful to untrained human users, but also can easily interface with semantic processing modules of larger agents that operate in real-time, such as tutoring systems.

The Task Matrix Framework for Platform-Independent Humanoid Programming

Programming humanoid robots is very difficult due to the significant demands imposed by managing balancing, locomotion, dynamics, and kinematic redundancy. Current research on humanoids tends to be dominated by these issues and generally focuses little on programming the robots to perform useful tasks. This paper discusses the task matrix, a framework that employs abstractions to allow roboticists to program humanoids at a high level and ignore the complex issues noted above. These abstractions also facilitate programming in a robot-independent manner, permitting software reuse. We examine the task matrix and show how it can be used to perform both simple and complex tasks on two simulated humanoid robots.