Brandon Rohrer

Motions or muscles? Some behavioral factors underlying robotic assistance of motor recovery

Robotics and related technologies have begun to realize their promise to improve the delivery of rehabilitation therapy. However, the mechanism by which they enhance recovery remains unclear. Ultimately, recovery depends on biology, yet the details of the recovery process remain largely unknown; a deeper understanding is important to accelerate refinements of robotic therapy or suggest new approaches. Fortunately, robots provide an excellent instrument platform from which to study recovery at the behavioral level. This article reviews some initial insights about the process of upper-limb behavioral recovery that have emerged from our work. Evidence to date suggests that the form of therapy may be more important than its intensity: muscle strengthening offers no advantage over movement training. Passive movement is insufficient; active participation is required. Progressive training based on measures of movement coordination yields substantially improved outcomes. Together these results indicate that movement coordination rather than muscle activation may be the most appropriate focus for robotic therapy.

Avoiding Spurious Submovement Decompositions II: A Scattershot Algorithm

Evidence for the existence of discrete submovements underlying continuous human movement has motivated many attempts to “extract” them. Although they produce visually convincing results, all of the methodologies that have been employed are prone to produce spurious decompositions. In previous work, a branch-and-bound algorithm for submovement extraction, capable of global nonlinear minimization, and hence, capable of avoiding spurious decompositions, was presented [Rohrer and Hogan (Biol Cybern 39:190–199, 2003)]. Here, we present a scattershot-type global nonlinear minimization algorithm that requires approximately four orders of magnitude less time to compute. A sensitivity analysis reveals that the scattershot algorithm can reliably detect changes in submovement parameters over time, e.g., over the course of neuromotor recovery.

Avoiding spurious submovement decompositions: a globally optimal algorithm

Abstract.Evidence for the existence of discrete submovements underlying continuous human movement has motivated many attempts to “extract” them. Although they produce visually convincing results, all of the methodologies that have been employed are prone to produce spurious decompositions. Examples of potential failures are given. A branch-and-bound algorithm for submovement extraction, capable of global nonlinear minimization (and hence capable of avoiding spurious decompositions), is developed and demonstrated.

Sparse imaging for fast electron microscopy

Scanning electron microscopes (SEMs) are used in neuroscience and materials science to image centimeters of sample area at nanometer scales. Since imaging rates are in large part SNR-limited, large collections can lead to weeks of around-the-clock imaging time. To increase data collection speed, we propose and demonstrate on an operational SEM a fast method to sparsely sample and reconstruct smooth images. To accurately localize the electron probe position at fast scan rates, we model the dynamics of the scan coils, and use the model to rapidly and accurately visit a randomly selected subset of pixel locations. Images are reconstructed from the undersampled data by compressed sensing inversion using image smoothness as a prior. We report image fidelity as a function of acquisition speed by comparing traditional raster to sparse imaging modes. Our approach is equally applicable to other domains of nanometer microscopy in which the time to position a probe is a limiting factor (e.g., atomic force microscopy), or in which excessive electron doses might otherwise alter the sample being observed (e.g., scanning transmission electron microscopy).

Model-Free Learning and Control in a Mobile Robot

A model-free, biologically-motivated learning and control algorithm called S-learning is described as implemented in an Surveyor SRV-1 mobile robot. S-learning demonstrated learning of robotic and environmental structure sufficient to allow it to achieve its goals (finding high- or low-contrast views in its environment). No modeling information about the task or calibration information about the robot’s actuators and sensors were used in S-learning’s planning. The ability of S-learning to make movement plans was completely dependent on experience it gained as it explored. Initially it had no experience and was forced to wander randomly. With increasing exposure to the task, S-learning achieved its goals with more nearly optimal paths. The fact that this approach is model-free implies that it may be applied to many other systems, perhaps even to systems of much greater complexity.

A learning and control approach based on the human neuromotor system

Current models of human motor learning and control typically employ continuous (or near continuous) movement commands and sensory information. However, research suggests that voluntary motor commands are issued in discrete-time submovements. There is also reasonable support for the hypothesis that human sensory experience is episodic as well. These facts have motivated the development of a learning algorithm that employs discrete-time sensory and motor control events, S-learning. We present this algorithm together with the results of simulated robot control. The results show that the learning that takes place is adaptive and is robust to a variety of conditions that many traditional controllers are not capable of handling, including random errors in the actuators and sensors, random transmission time delays, hard nonlinearities, time varying system behavior, and unknown structure of system dynamics. The performance of S-learning suggests that it may be an appropriate high-level control scheme for complex robotic systems, including walking, cooperative manipulation, and humanoid robots

S-Learning: A Biomimetic Algorithm for Learning, Memory, and Control in Robots

S-learning is a sequence-based learning algorithm patterned on human motor behavior. Discrete-time and quantized sensory information is amassed in real-time to form a dynamic model of the system being controlled and its environment. No explicit model is provided a priori, nor any hint about what the structure of the model might be. As the core of a Brain-Emulating Cognition and Control Architecture (BECCA), S-Learning provides a mechanism for human-inspired learning, memory, and control in machines. In a simulation of a point-to-point reaching task, S-Learning demonstrates several attributes of human motor behavior, including learning through exploration and task transfer.

In-Sole MEMS Pressure Sensing for a LowerExtremity Exoskeleton

The control system for the Berkeley lower extremity exoskeleton (BLEEX) requires ground contact pressure information to operate safely and effectively. Commercially available in-sole sensors do not have sufficient bandwidth, accuracy and reliability for such a system. We have designed and prototyped an in-sole ground contact sensor that uses MEMS pressure transducers placed in an array of hermetically sealed cavities. This system provides a robust method to monitor ground contact pressures

MEMS-based bubble pressure sensor for prosthetic socket interface pressure measurement

The ability to chronically monitor pressure at the prosthetic socket/residual limb interface could provide important data to the research and clinical communities. With this application in mind, we describe a novel type of sensor which consists of a MEMS pressure sensor and custom electronics packaged in a fluid-filled bubble. The sensor is characterized and compared to two commercially-available technologies. The bubble sensor has excellent drift performance and good sensing resolution. It exhibits hysteresis which may be due to the silicone that the sensor is molded in. To reduce hysteresis, it may be advisable to place the sensor between the liner and the socket rather molding directly into the liner.

S-Learning: A Model-Free, Case-Based Algorithm for Robot Learning and Control

A model-free, case-based learning and control algorithm called S-learning is described as implemented in a simulation of a light-seeking mobile robot. S-learning demonstrated learning of robotic and environmental structure sufficient to allow it to achieve its goal (reaching a light source). No modeling information about the task or calibration information about the robots actuators and sensors were used in S-learnings planning. The ability of S-learning to make movement plans was completely dependent on experience it gained as it explored. Initially it had no experience and was forced to wander randomly. With increasing exposure to the task, S-learning achieved its goal with more nearly optimal paths. The fact that this approach is model-free and case-based implies that it may be applied to many other systems, perhaps even to systems of much greater complexity.