Search

Real-world autonomous systems are often controlled using conventional model-based control methods. But if accurate models of a system are not available, these methods may be unsuitable. For many safety-critical systems, such as robotic systems, a model of the system and a control strategy may be learned using data. When applying learning to safety-critical systems, guaranteeing safety during learning as well as testing/deployment is paramount. A variety of different approaches for ensuring safety exists, but the published works are cluttered and there are few reviews that compare the latest approaches. This paper reviews two promising approaches on guaranteeing safety for learning-based robust control of uncertain dynamical systems, which are based on control barrier functions and control Lyapunov functions. While control barrier functions provide an option to incorporate safety in terms of constraint satisfaction, control Lyapunov functions are used to define safety in terms of stability. This review categorises learning-based methods that use control barrier functions and control Lyapunov functions into three groups, namely reinforcement learning, online and offline supervised learning. Finally, the paper presents a discussion of the suitability of the different methods for different applications.

In this paper, we propose a deep learning instance segmentation framework for particle extraction of microscopic images that aims at calculating planktonic species distribution and concentration in-situ. The framework comprises three essential functional tasks on in-situ time-series images collected from an autonomous underwater vehicle: 1) manual labeling of the captured images, 2) object localization, segmentation, and identification, and 3) class distribution and planktonic organisms concentration calculation. Our proposed framework is based on the mask R-CNN architecture provided by the Detectron2 library developed by Facebook Artificial Intelligence Research (FAIR) for instance segmentation. Due to its modular design, we compare the performance of different networks by alternating the backbone sub-network in order to choose the most suitable architecture for the task of instance and semantic segmentation. We compile a custom annotated dataset from planktonic time-series images and train the different models over this dataset to perform the instance semantic segmentation. Evaluation results of the proposed framework, utilizing the best performing deep learning architecture along with the new annotated dataset, show better performance in terms of speed and accuracy of both in-situ segmentation and classification compared to traditional segmentation methods. In addition, we observe a significant improvement in the object classification quality when we train the model over our newly annotated dataset instead of training it over the dataset generated from the traditional methods. The inferred data from our novel instance segmentation framework, which provides the particle class distribution and concentration, can then be used to assist in constructing a dynamic probability density map of planktonic communities dispersion and abundance.

With the complex structure of planktonic species and an immense amount of data captured from autonomous underwater vehicles (AUVs), a large burden is placed on the domain experts for plankton taxa labeling. At the same time, the most prominent machine learning (ML) methods for classification rely heavily on a massive amount of labeled datasets to create and train neural network classifier models that perform their tasks accurately. Active Learning (AL) is an ML paradigm that reduces this manual effort by proposing algorithms that support the construction of the training datasets, thus enlarging the sets while minimizing human involvement. To build the training set, AL methods apply heuristics to select a subset of images, i.e., samples, from the entire data. The selected samples that capture the common statistical patterns or feature space are likely to include all the information needed for the training and the learning processes. In addition, the algorithm should prioritize samples that are likely belonging to multiple classes, i.e., having close inter-class boundaries, and might lead to model confusion. Many of the current AL approaches fail to incorporate both types of samples representing the statistical pattern and the samples in which the particular machine learning model is uncertain about. In this paper, we extend our framework which addresses these challenges with an augmentation module to increase the robustness of the model and ensure its adaptability to the planktonic domain. We compare the framework with existing hybrid AL techniques and test an adaption of our extended framework on the planktonic domain. The empirical results from the experiments exerted in this paper confirm higher accuracy achieved by the new extended framework.

Autonomous cyber-physical systems, CPS, in dynamic environments must work impeccably. The cyber-physical systems must handle tasks consistently and trustworthily, i.e., with a robust behavior. Robust systems, in general, require making valid and solid decisions using one or a combination of robust reasoning strategies, algorithms, and robustness analysis. However, in dynamic environments, data can be incomplete, skewed, contradictory, and redundant impacting the reasoning. Basing decisions on these data can lead to inconsistent, irrational, and unreasonable cyber-physical systems’ movements, adversely impacting the system’s reliability and integrity. This paper presents the assessment of robust reasoning for autonomous cyber-physical systems in dynamic environments. In this work, robust reasoning is considered as 1) the capability of drawing conclusions with available data by applying classical and non-classical reasoning strategies and algorithms and 2) act and react robustly and safely in dynamic environments by employing robustness analysis to provide options on possible actions and evaluate alternative decisions. The result of the research shows that different common existing strategies, algorithms and analyses can be provided together with a comparison of their applicabilities, benefits, and drawbacks in the context of cyber-physical systems operating in dynamically changing environments. The conclusion is that robust reasoning in cyber-physical systems can handle dynamic environments. Moreover, combining these strategies and algorithms with robustness analysis can support achieving robust behavior in autonomous cyber-physical systems while operating in dynamically changing environments.

Planktonic organisms form the principal source for consumers on higher trophic levels in the food chain. Studying their community dispersion is vital to our understanding of the planet's ecological systems. With the recent technological advancements in imaging systems, capturing images of planktons in-situ is made possible by embedding mobile underwater robots with sophisticated camera systems and computing power that implement deep machine learning approaches. Efforts of applying deep learning methods to plankton imaging systems have been limited to classification, while detection and segmentation has been left to traditional methods in this context. There is a variety of publicly available datasets made suited for planktonic species classification. These datasets consist of images of individual species. Thus, they do not represent the actual environment, which is usually given by a scene representation more suited for object localization, detection and semantic segmentation. In this paper we propose a novel custom dataset [1] from planktonic images captured in-situ in a lab environment suited for supervised learning of object detection and instance segmentation. The data is tested in experiments using the state-of-the-art deep learning visual recognition method of Mask R-CNN. The experiment results show the potential of this method and create a baseline analysis module for real-time in-situ image processing. We provide a comparison of how the method is performing when trained on automatically processed and annotated images from existing segmentation frameworks using traditional methods. This comparison illustrates the importance of utilizing proper data and the potential for success if provided11All results, code and metrics used for the experiments are provided in: https://github.com/AILARON/Segmentation.

Planktons are vital to our planet's ecosystems. Mapping and monitoring planktonic-distribution in the ocean is of great importance in understanding these ecosystems as well as gaining insights on the general health of our planet. The task of identifying different planktonic species and their concentrations is a critical yet challenging part of this endeavor. In this paper, we explore the utilization of one-shot classifiers on in-situ captured planktonic images. The similarity-based classification method used by the one-shot classification algorithms gives rise to various research opportunities outside of the typical classification paradigm. Exploring the development of a proficient plankton image classification framework that can determine how different or how similar two plankton specimens are, can inform us if they belong to the same species, if they are different variations of the same species or if we have encountered a completely unseen classes of plankton. This similarity is valuable information that assists in further mapping the diversity of the planktonic species. We further extend the Siamese one-shot classifier with few-shot classifier to improve the model performance. Empirical evaluations of the new extension yield promising results.

Rapid assessment and enhanced knowledge of plankton communities and their structure in the productive upper water column is of crucial importance to understand the impact of the changing climate on upper ocean processes. Enabling persistent and systematic observation by coupling the ongoing revolution in robotics and automation with artificial intelligence methods will improve accuracy of predictions, reduce measurement uncertainty and accelerate methodological sampling. Further, progress in real-time robotic visual sensing and machine learning have enabled high-resolution space-time imaging, analysis and interpretation. We describe a novel mobile robotic tool for characterizing upper water-column biota, enabling intelligent onboard sampling to target specific mesoplankton taxa. Such a tool will accelerate the time consuming process in asking ``who is there'' and aid the advancement of oceanographic observation.

This paper introduces a new constraint domain for reasoning about data with uncertainty. It extends convex modeling with the notion of p-box to gain additional quantifiable information on the data whereabouts. Unlike existing approaches, the p-box envelops an unknown probability instead of approximating its representation. The p-box bounds are uniform cumulative distribution functions (cdf) in order to employ linear computations in the probabilistic domain. The reasoning by means of p-box cdf-intervals is an interval computation which is exerted on the real domain then it is projected onto the cdf domain. This operation conveys additional knowledge represented by the obtained probabilistic bounds. The empirical evaluation of our implementation shows that, with minimal overhead, the output solution set realizes a full enclosure of the data along with tighter bounds on its probabilistic distributions.

International audience; Interval coefficients have been introduced in OR and CP to specify uncertain data in order to provide reliable solutions to convex models. The output is generally a solution set, guaranteed to contain all solutions possible under any realization of the data. This set can be too large to be meaningful. Furthermore, each solution has equal uncertainty weight, thus does not reflect any possible degree of knowledge about the data. To overcome these problems we propose to extend the notion of interval coefficient by introducing a second dimension to each interval bound. Each bound is now specified by its data value and its degree of knowledge. This is formalized using the cumulative distribution function of the data set. We define the formal framework of constraint reasoning over this cdf-intervals. The main contribution of this paper concerns the formal definition of a new interval arithmetic and its implementation. Promising results on problem instances demonstrate the approach.