Danesh Tarapore

Robots that can adapt like animals

Robots have transformed many industries, most notably manufacturing, and have the power to deliver tremendous benefits to society, such as in search and rescue, disaster response, health care and transportation. They are also invaluable tools for scientific exploration in environments inaccessible to humans, from distant planets to deep oceans. A major obstacle to their widespread adoption in more complex environments outside factories is their fragility. Whereas animals can quickly adapt to injuries, current robots cannot ‘think outside the box’ to find a compensatory behaviour when they are damaged: they are limited to their pre-specified self-sensing abilities, can diagnose only anticipated failure modes, and require a pre-programmed contingency plan for every type of potential damage, an impracticality for complex robots. A promising approach to reducing robot fragility involves having robots learn appropriate behaviours in response to damage, but current techniques are slow even with small, constrained search spaces. Here we introduce an intelligent trial-and-error algorithm that allows robots to adapt to damage in less than two minutes in large search spaces without requiring self-diagnosis or pre-specified contingency plans. Before the robot is deployed, it uses a novel technique to create a detailed map of the space of high-performing behaviours. This map represents the robot’s prior knowledge about what behaviours it can perform and their value. When the robot is damaged, it uses this prior knowledge to guide a trial-and-error learning algorithm that conducts intelligent experiments to rapidly discover a behaviour that compensates for the damage. Experiments reveal successful adaptations for a legged robot injured in five different ways, including damaged, broken, and missing legs, and for a robotic arm with joints broken in 14 different ways. This new algorithm will enable more robust, effective, autonomous robots, and may shed light on the principles that animals use to adapt to injury.

Neural networks as mechanisms to regulate division of labor

In social insects, workers perform a multitude of tasks, such as foraging, nest construction, and brood rearing, without central control of how work is allocated among individuals. It has been suggested that workers choose a task by responding to stimuli gathered from the environment. Response-threshold models assume that individuals in a colony vary in the stimulus intensity (response threshold) at which they begin to perform the corresponding task. Here we highlight the limitations of these models with respect to colony performance in task allocation. First, we show with analysis and quantitative simulations that the deterministic response-threshold model constrains the workers’ behavioral flexibility under some stimulus conditions. Next, we show that the probabilistic response-threshold model fails to explain precise colony responses to varying stimuli. Both of these limitations would be detrimental to colony performance when dynamic and precise task allocation is needed. To address these problems, we propose extensions of the response-threshold model by adding variables that weigh stimuli. We test the extended response-threshold model in a foraging scenario and show in simulations that it results in an efficient task allocation. Finally, we show that response-threshold models can be formulated as artificial neural networks, which consequently provide a comprehensive framework for modeling task allocation in social insects.

Task-dependent influence of genetic architecture and mating frequency on division of labour in social insect societies

Division of labour is one of the most prominent features of social insects. The efficient allocation of individuals to different tasks requires dynamic adjustment in response to environmental perturbations. Theoretical models suggest that the colony-level flexibility in responding to external changes and internal perturbation may depend on the within-colony genetic diversity, which is affected by the number of breeding individuals. However, these models have not considered the genetic architecture underlying the propensity of workers to perform the various tasks. Here, we investigated how both within-colony genetic variability (stemming from variation in the number of matings by queens) and the number of genes influencing the stimulus (threshold) for a given task at which workers begin to perform that task jointly influence task allocation efficiency. We used a numerical agent-based model to investigate the situation where workers had to perform either a regulatory task or a foraging task. One hundred generations of artificial selection in populations consisting of 500 colonies revealed that an increased number of matings always improved colony performance, whatever the number of loci encoding the thresholds of the regulatory and foraging tasks. However, the beneficial effect of additional matings was particularly important when the genetic architecture of queens comprised one or a few genes for the foraging task’s threshold. By contrast, a higher number of genes encoding the foraging task reduced colony performance with the detrimental effect being stronger when queens had mated with several males. Finally, the number of genes encoding the threshold for the regulatory task only had a minor effect on colony performance. Overall, our numerical experiments support the importance of mating frequency on efficiency of division of labour and also reveal complex interactions between the number of matings and genetic architecture.

Quantifying patterns of agent-environment interaction

This article explores the assumption that a deeper (quantitative) understanding of the information-theoretic implications of sensory-motor coordination can help endow robots not only with better sensory morphologies, but also with better exploration strategies. Speciflcally, we investigate by means of statistical and informationtheoretic measures, to what extent sensory-motor coordinated activity can generate and structure information in the sensory channels of a simulated agent interacting with its surrounding environment. The results show how the usage of correlation, entropy, and mutual information can be employed (a) to segment an observed behavior into distinct behavioral states, (b) to analyze the informational relationship between the difierent components of the sensory-motor apparatus, and (c) to identify patterns (or flngerprints) in the sensory-motor interaction between the agent and its local environment.

To err is robotic, to tolerate immunological: fault detection in multirobot systems

Fault detection and fault tolerance represent two of the most important and largely unsolved issues in the field of multirobot systems (MRS). Efficient, long-term operation requires an accurate, timely detection, and accommodation of abnormally behaving robots. Most existing approaches to fault-tolerance prescribe a characterization of normal robot behaviours, and train a model to recognize these behaviours. Behaviours unrecognized by the model are consequently labelled abnormal or faulty. MRS employing these models do not transition well to scenarios involving temporal variations in behaviour (e.g., online learning of new behaviours, or in response to environment perturbations). The vertebrate immune system is a complex distributed system capable of learning to tolerate the organisms tissues even when they change during puberty or metamorphosis, and to mount specific responses to invading pathogens, all without the need of a genetically hardwired characterization of normality. We present a generic abnormality detection approach based on a model of the adaptive immune system, and evaluate the approach in a swarm of robots. Our results reveal the robust detection of abnormal robots simulating common electro-mechanical and software faults, irrespective of temporal changes in swarm behaviour. Abnormality detection is shown to be scalable in terms of the number of robots in the swarm, and in terms of the size of the behaviour classification space.

Generic, scalable and decentralized fault detection for robot swarms

Robot swarms are large-scale multirobot systems with decentralized control which means that each robot acts based only on local perception and on local coordination with neighboring robots. The decentralized approach to control confers number of potential benefits. In particular, inherent scalability and robustness are often highlighted as key distinguishing features of robot swarms compared with systems that rely on traditional approaches to multirobot coordination. It has, however, been shown that swarm robotics systems are not always fault tolerant. To realize the robustness potential of robot swarms, it is thus essential to give systems the capacity to actively detect and accommodate faults. In this paper, we present a generic fault-detection system for robot swarms. We show how robots with limited and imperfect sensing capabilities are able to observe and classify the behavior of one another. In order to achieve this, the underlying classifier is an immune system-inspired algorithm that learns to distinguish between normal behavior and abnormal behavior online. Through a series of experiments, we systematically assess the performance of our approach in a detailed simulation environment. In particular, we analyze our system’s capacity to correctly detect robots with faults, false positive rates, performance in a foraging task in which each robot exhibits a composite behavior, and performance under perturbations of the task environment. Results show that our generic fault-detection system is robust, that it is able to detect faults in a timely manner, and that it achieves a low false positive rate. The developed fault-detection system has the potential to enable long-term autonomy for robust multirobot systems, thus increasing the usefulness of robots for a diverse repertoire of upcoming applications in the area of distributed intelligent automation.

Abnormality Detection in Robots Exhibiting Composite Swarm Behaviours

Fault detection is one of the most prominent challenges in the field of multirobot systems (MRS). Most existing faulttolerant systems prescribe a characterisation of normal behaviours (fault-free behaviours), and train a model to recognise them. Behaviours not recognised by the model are labelled abnormal. MRS employing these models do not transition well to scenarios involving gradual changes in normal behaviour. In such scenarios, existing fault-detection systems may not be applicable, or may incur potentially costly false positive detections. We propose to address this challenging problem by taking inspiration from the regulation of tolerance and (auto)immunity in the adaptive immune system. We deploy an immune system-based fault-detection approach to detect abnormalities in heterogeneously behaving robots. Results of extensive simulation-based experiments demonstrate that a distributed MRS can correctly tolerate delayed propagation of different normal behaviours in the collective, at low false-positive rates. Furthermore, the fault-detection system is able to reliably detect robots performing different faultsimulating behaviours.

Towards Fault Diagnosis in Robot Swarms: An Online Behaviour Characterisation Approach

Although robustness has been cited as an inherent advantage of swarm robotics systems, it has been shown that this is not always the case. Fault diagnosis will be critical for future swarm robotics systems if they are to retain their advantages (robustness, flexibility and scalability). In this paper, existing work on fault detection is used as a foundation to propose a novel approach for fault diagnosis in swarms based on a behavioural feature vector approach. Initial results show that behavioural feature vectors can be used to reliably diagnose common electro-mechanical fault types in most cases tested.

Optimizing the crossregulation model for scalable abnormality detection

The engineering of fault-detection systems for multirobot systems (MRS) is a well-studied problem (e.g, Christensen et al. (2009)). Most fault-detection models are built on the assumption that normal behavior is known, and can be characterized in advance. The models are trained to recognize predefined normal behaviors, and behaviors not recognized are labeled abnormal. While such an approach does provide some interesting results of robust fault detection and fault tolerance, they may not be applicable when normal behavior can change as a result of unforeseen environmental conditions and online learning for instance. Furthermore, prior information required to characterize normal behaviors, may not always be available.