Alan Diamond

ECCE1: The first of a series of anthropomimetic musculoskeletal upper torsos

The human body was not designed by engineers and the way in which it is built poses enormous control problems. Its complexity challenges the ability of classical control theory to explain human movement as well as the development of human motor skills. It is our working hypothesis that the engineering paradigm for building robots places severe limitations on the kinds of interactions such robots can engage in, on the knowledge they can acquire of their environment, and therefore on the nature of their cognitive engagement with the environment. This paper describes the design of an anthropomimetic humanoid upper torso, ECCE1, built in the context of the ECCEROBOT project. The goal of the project is to use this platform to test hypotheses about human motion as well as to compare its performance with that of humans, whether at the mechanical, behavioural or cognitive level.

Toward anthropomimetic robotics: Development, simulation, and control of a musculoskeletal torso

Anthropomimetic robotics differs from conventional approaches by capitalizing on the replication of the inner structures of the human body, such as muscles, tendons, bones, and joints. Here we present our results of more than three years of research in constructing, simulating, and, most importantly, controlling anthropomimetic robots. We manufactured four physical torsos, each more complex than its predecessor, and developed the tools required to simulate their behavior. Furthermore, six different control approaches, inspired by classical control theory, machine learning, and neuroscience, were developed and evaluated via these simulations or in small-scale setups. While the obtained results are encouraging, we are aware that we have barely exploited the potential of the anthropomimetic design so far. But, with the tools developed, we are confident that this novel approach will contribute to our understanding of morphological computation and human motor control in the future.

Towards anthropomimetic robotics: Development, simulation, and control of a musculoskeletal torso

Abstract Anthropomimetic robotics differs from conventional approaches by capitalizing on the replication of the inner structures of the human body, such as muscles, tendons, bones, and joints. Here we present our results of more than three years of research in constructing, simulating, and, most importantly, controlling anthropomimetic robots. We manufactured four physical torsos, each more complex than its predecessor, and developed the tools required to simulate their behavior. Furthermore, six different control approaches, inspired by classical control theory, machine learning, and neuroscience, have been developed and evaluated via these simulations or in small-scale setups. While the obtained results are encouraging, we are aware that we have barely exploited the potential of the anthropomimetic design so far. But, with the tools developed, we are confident that this novel approach will contribute to our understanding of morphological computation and human motor control in the future.

Comparing neuromorphic solutions in action: implementing a bio-inspired solution to a benchmark classification task on three parallel-computing platforms

Neuromorphic computing employs models of neuronal circuits to solve computing problems. Neuromorphic hardware systems are now becoming more widely available and “neuromorphic algorithms” are being developed. As they are maturing toward deployment in general research environments, it becomes important to assess and compare them in the context of the applications they are meant to solve. This should encompass not just task performance, but also ease of implementation, speed of processing, scalability, and power efficiency. Here, we report our practical experience of implementing a bio-inspired, spiking network for multivariate classification on three different platforms: the hybrid digital/analog Spikey system, the digital spike-based SpiNNaker system, and GeNN, a meta-compiler for parallel GPU hardware. We assess performance using a standard hand-written digit classification task. We found that whilst a different implementation approach was required for each platform, classification performances remained in line. This suggests that all three implementations were able to exercise the models ability to solve the task rather than exposing inherent platform limits, although differences emerged when capacity was approached. With respect to execution speed and power consumption, we found that for each platform a large fraction of the computing time was spent outside of the neuromorphic device, on the host machine. Time was spent in a range of combinations of preparing the model, encoding suitable input spiking data, shifting data, and decoding spike-encoded results. This is also where a large proportion of the total power was consumed, most markedly for the SpiNNaker and Spikey systems. We conclude that the simulation efficiency advantage of the assessed specialized hardware systems is easily lost in excessive host-device communication, or non-neuronal parts of the computation. These results emphasize the need to optimize the host-device communication architecture for scalability, maximum throughput, and minimum latency. Moreover, our results indicate that special attention should be paid to minimize host-device communication when designing and implementing networks for efficient neuromorphic computing.

Anthropomimetic Robots: Concept, Construction and Modelling

An anthropomimetic robot is one that closely copies the mechanics of the human body by having a human-like jointed skeleton moved by compliant muscle-like actuators. This paper describes the progress achieved in building anthropomimetic torsos in two projects, CRONOS and ECCEROBOT. In each, the bones were hand-moulded in a thermoplastic and the muscles were implemented by DC motors shortening and extending elastic tendons. Anthropomimetic robots differ from conventionally engineered robots by having complex joints and compliant tendon driven actuation that can cross more than one joint. Taken together, these characteristics make the robots unsuitable for control by standard methods, and so the ability to model them is important for developing heuristic methods of control and also for providing forward models. The robots were modelled using physics-based techniques which enable the study of the generation of movements and also of interactions with arbitrary objects. The lightweight and compliant structure of the robots was found to be safe for human proximity and contact.

Using the Microsoft Kinect to model the environment of an anthropomimetic robot

The control of compliantly actuated anthropomimetic robots with complex and multiarticular joints, such as those developed within the ECCEROBOT project, is extremely challenging. We are approaching the problem by using a physics engine to run a highly detailed simulation of such a robot’s structure and dynamic behaviour, and then searching for sequences of motor activations that will achieve particular goals. This requires the simulated robot to be situated accurately in a physics-based representation of its environment which includes the object with which it is to interact. In this paper we present our environmental sensing and modelling scheme which uses data from a single headmounted Kinect sensor to provide and locate the environmental model, and to identify and locate the target object accurately in the presence of significant motion blur.

Classifying chemical sensor data using GPU-accelerated bio-mimetic neuronal networks based on the insect olfactory system

Chemosensing “e-nose” technology has great potential applications in everyday life, ranging from drug detection to food quality assessment and even the diagnosis of illness. However, odour detection and classification remains a highly challenging domain, characterized by high dimensionality, unknown organization of the vast “odourant space” of volatile chemicals and complex, turbulent odour plumes. To compound these difficulties, current sensor technology continues to exhibit distinct shortcomings in speed, sensitivity, selectivity, recovery, and drift avoidance. In the research reported here we turn to a range of recent neuronal models [1-6] that were developed to describe the insect olfactory system, and have been shown to perform well across a range of classic, static classification tasks such as the MNIST handwritten digit set and the Sigma-Aldrich scent database. The insect olfactory system has been extensively studied and has been shown to be both fast and highly effective in complex natural conditions despite its limited size and complexity (when compared to the mammalian system) [2]. Insects such as moths, honey bees, locusts and fruit flies are capable of odour detection and classification tasks well beyond the abilities of current e-nose technology and machine learning algorithms [6]. We present results of applying an insect-inspired approach to the design of a learning spiking neural network that receives synchronized time series data from up to 12 metal-oxide based gas sensors, comprising an optimised [7] combination of classical doped tin oxide and novel zeolite-coated chromium titanium oxide sensors. We have collected sample data sets for classification tasks ranging from “easy” (single chemical identification presented under laboratory conditions) through to “hard” (identification of indicators of infectious diseases in breath samples taken from patients). To address slow sensor response we consider a range of transient-based processing before applying self-organisation techniques to most effectively locate “virtual receptors” (VR) in sensor space. We look to address decorrelation (separation in feature space) and the supervised association of responses with rewards by using correlates of the insect antennal lobe (AL) and mushroom body (MB) structures whilst applying reward-based spike-timing dependent plasticity mechanisms. Classification accuracy is compared with support vector machine (SVM) learning which we also look to match for speed through the use of GPU accelerated neural simulation [5] via the NVidia CUDA TM -based GeNN platform (http://sourceforge.net/projects/genn/).