The last half decade has seen a steep rise in the number of contributions on safe learning methods for real-world robotic deployments from both the control and reinforcement learning communities. This article provides a concise but holistic review of the recent advances made in using machine learning to achieve safe decision-making under uncertainties, with a focus on unifying the language and frameworks used in control theory and reinforcement learning research. It includes learning-based control approaches that safely improve performance by learning the uncertain dynamics, reinforcement learning approaches that encourage safety or robustness, and methods that can formally certify the safety of a learned control policy. As data- and learning-based robot control methods continue to gain traction, researchers must understand when and how to best leverage them in real-world scenarios where safety is imperative, such as when operating in close proximity to humans. We highlight some of the open challenges that will drive the field of robot learning in the coming years, and emphasize the need for realistic physics-based benchmarks to facilitate fair comparisons between control and reinforcement learning approaches. Expected final online publication date for the Annual Review of Control, Robotics, and Autonomous Systems, Volume 5 is May 2022. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

The development of the highest resolution, large-area, active-matrix, flat-panel imager (AMFPI) thus far reported is described. This imager is based on a 97 jtm pixel pitch array with each pixel comprising a single a-Si:H TFT coupled to a discrete a-Si:H n-i-p photodiode. While the initial configuration chosen for fabrication is a 2048x2048 pixel array, a larger monolithic array format of 3072x4096 pixels is also permitted by the design. When coupled to an overlying scintillator such as a phosphor screen or CsI:Tl, the array allows indirect detection of incident radiation. The array is operated in conjunction with a recently completed electronic acquisition system featuring asynchronous operation, a large addressing range, fast analog signal extraction and digitization, and 16-bit digitization. This imager, whose empirical characterization will be reported in a subsequent paper, was developed as an engineering prototype to allow investigation of the performance limits of the most aggressive array designs permitted by present active-matrix technology. The development of this new imager builds upon knowledge acquired from the iterative design, fabrication, and quantitative evaluation of earlier engineering prototypes based on a series of 127 im pitch arrays. This paper summarizes the general program of research leading to this new device and puts this in the context of world-wide developments in indirect and direct detection AMFPI technology. Some limitations of present AMFPI technology are described, and possible solutions are discussed.Specifically, the incorporation of multiplexers based on poly-crystalline silicon circuitry into the array design, to facilitate very high resolution imagers, are proposed. In addition, strategies to significantly improve AMFPI performance at very low exposures, such as those commonly encountered in fluoroscopy, involving the reduction of additive noise (such as through lower preamplifier noise) and the enhancement of system gain (such as through the use of lead iodide) are discussed and initial calculations illustrating potential levels ofperformance are presented.

We have developed workflows to align 3D magnetic resonance histology (MRH) of the mouse brain with light sheet microscopy (LSM) and 3D delineations of the same specimen. We start with MRH of the brain in the skull with gradient echo and diffusion tensor imaging (DTI) at 15 μm isotropic resolution which is ~ 1,000 times higher than that of most preclinical MRI. Connectomes are generated with superresolution tract density images of ~5 μm. Brains are cleared, stained for selected proteins, and imaged by LSM at 1.8 μm/pixel. LSM data are registered into the reference MRH space with labels derived from the ABA common coordinate framework. The result is a  hi gh- d imensional  i ntegrated  v olum e  with  r egistration ( HiDiver ) with alignment precision better than 50 µm. Throughput is sufficiently high that HiDiver is being used in quantitative studies of the impact of gene variants and aging on mouse brain cytoarchitecture and connectomics.

We have developed new imaging and computational workflows to produce accurately aligned multimodal 3D images of the mouse brain that exploit high resolution magnetic resonance histology (MRH) and light sheet microscopy (LSM) with fully rendered 3D reference delineations of brain structures. The suite of methods starts with the acquisition of geometrically accurate (in-skull) brain MRIs using multi-gradient echo (MGRE) and new diffusion tensor imaging (DTI) at an isotropic spatial resolution of 15 um. Whole brain connectomes are generated using over 100 diffusion weighted images acquired with gradients at uniformly spaced angles. Track density images are generated at a super-resolution of 5 um. Brains are dissected from the cranium, cleared with SHIELD, stained by immunohistochemistry, and imaged by LSM at 1.8 um/pixel. LSM channels are registered into the reference MRH space along with the Allen Brain Atlas (ABA) Common Coordinate Framework version 3 (CCFv3). The result is a high-dimensional integrated volume with registration (HiDiver) that has a global alignment accuracy of 10-50 um. HiDiver enables 3D quantitative and global analyses of cells, circuits, connectomes, and CNS regions of interest (ROIs). Throughput is sufficiently high that HiDiver is now being used in comprehensive quantitative studies of the impact of gene variants and aging on rodent brain cytoarchitecture.

The last half-decade has seen a steep rise in the number of contributions on safe learning methods for real-world robotic deployments from both the control and reinforcement learning communities. This article provides a concise but holistic review of the recent advances made in using machine learning to achieve safe decision making under uncertainties, with a focus on unifying the language and frameworks used in control theory and reinforcement learning research. Our review includes: learning-based control approaches that safely improve performance by learning the uncertain dynamics, reinforcement learning approaches that encourage safety or robustness, and methods that can formally certify the safety of a learned control policy. As dataand learning-based robot control methods continue to gain traction, researchers must understand when and how to best leverage them in real-world scenarios where safety is imperative, such as when operating in close proximity to humans. We highlight some of the open challenges that will drive the field of robot learning in the coming years, and emphasize the need for realistic physics-based benchmarks to facilitate fair comparisons between control and reinforcement learning approaches.