Developers and academics have grown increasingly interested in developing “trustworthy” artificial intelligence (AI). However, this aim is difficult to achieve in practice, especially given trust and trustworthiness are complex, multifaceted concepts that cannot be completely guaranteed nor built entirely into an AI system. We have drawn on the breadth of trust-related literature across multiple disciplines and fields to synthesize knowledge pertaining to interpersonal trust, trust in automation, and risk and trust. Based on this review we have (re)conceptualized trustworthiness in practice as being both (a) perceptual, meaning that a user assesses whether, when, and to what extent AI model output is trustworthy, even if it has been developed in adherence to AI trustworthiness standards, and (b) context-dependent, meaning that a user's perceived trustworthiness and use of an AI model can vary based on the specifics of their situation (e.g., time-pressures for decision-making, high-stakes decisions). We provide our reconceptualization to nuance how trustworthiness is thought about, studied, and evaluated by the AI community in ways that are more aligned with past theoretical research.

Knowledge of three-dimensional (3D) cloud structures is critical for a wide range of weather and climate studies. We recently introduced satellite-based cloud vertical cross-section products for aviation weather applications. They are based on newly-processed 3D cloud data that combines multiple satellite cloud products obtained from the NOAA Enterprise Cloud Algorithm Suite and additional environmental datasets. Leveraging these research efforts and being mindful of the current and future states of global cloud data development activities, this work is now expanded to generate global 3D cloud fields by applying cloud retrieval algorithms to multiple Low Earth Orbit (LEO) and Geostationary (GEO) satellite sensors. We also incorporate enhanced cloud properties into the global framework by adopting Artificial intelligence (AI)/Machine Learning (ML) approaches. This paper presents our current accomplishments and continuing efforts for further improved satellite cloud data.

Geosciences is a field of great societal relevance that requires solutions to several urgent problems facing our humanity and the planet. As geosciences enters the era of big data, machine learning (ML)-that has been widely successful in commercial domains-offers immense potential to contribute to problems in geosciences. However, geoscience applications introduce novel challenges for ML due to combinations of geoscience properties encountered in every problem, requiring novel research in machine learning. This article introduces researchers in the machine learning (ML) community to these challenges offered by geoscience problems and the opportunities that exist for advancing both machine learning and geosciences. We first highlight typical sources of geoscience data and describe their common properties. We then describe some of the common categories of geoscience problems where machine learning can play a role, discussing the challenges faced by existing ML methods and opportunities for novel ML research. We conclude by discussing some of the cross-cutting research themes in machine learning that are applicable across several geoscience problems, and the importance of a deep collaboration between machine learning and geosciences for synergistic advancements in both disciplines.

In this paper, we consider the use of structure learning methods for probabilistic graphical models to identify statistical dependencies in high-dimensional physical processes. Such processes are often synthetically characterized using PDEs (partial differential equations) and are observed in a variety of natural phenomena. In this paper, we present ACLIME-ADMM, an efficient two-step algorithm for adaptive structure learning, which decides a suitable edge specific threshold in a data-driven statistically rigorous manner. Both steps of our algorithm use (inexact) ADMM to solve suitable linear programs, and all iterations can be done in closed form in an efficient block parallel manner. We compare ACLIME-ADMM with baselines on both synthetic data simulated by PDEs that model advection-diffusion processes, and real data of daily global geopotential heights to study information flow in the atmosphere. ACLIME-ADMM is shown to be efficient, stable, and competitive, usually better than the baselines especially on difficult problems. On real data, ACLIME-ADMM recovers the underlying structure of global atmospheric circulation, including switches in wind directions at the equator and tropics entirely from the data.

Causal discovery algorithms have been used to identify potential cause-effect relationships from observational data for decades. Recently more applications are emerging, for example in climate science, that extend over large spatial domains and require temporal models. This paper first reviews how the causal discovery problem can be set up for such spatiotemporal problems using constraint-based structure learning, then discusses pitfalls we encountered and some solutions we developed. In particular, we consider how to handle temporal and spatial boundaries (which often result in causal sufficiency violations) and discuss the effects of temporal resolution and grid irregularities on the resulting model.

This paper investigates the robustness of underconstrained cable-driven robots to external disturbance wrenches (force/moment combinations). Two cases are considered: impulsive disturbance wrenches and static (constant) disturbance wrenches. In the analysis of these cases a mapping to an intermediate space is introduced, which allows a restatement of Gauss' principle of least constraint in a simpler form. Two measures are developed: the impulsive disturbance robustness measure and the static disturbance robustness measure. It is then shown that these two measures can be expressed as a single disturbance robustness measure, R. This single measure enables underconstrained cable robots to be designed for maximum robustness to both impulsive and static disturbances

This paper presents a method for analytically generating the boundaries of the wrench-feasible workspace (WFW) for cable robots. This method uses the available net wrench set, which is the set of all wrenches that a cable robot can apply to its surroundings without violating tension limits in the cables. The geometric properties of this set permit calculation of the boundaries of the WFW for planar, spatial, and point-mass cable robots. Complete analytical expressions for the WFW boundaries are detailed for a planar cable robot and a spatial point-mass cable robot. The analytically determined boundaries are verified by comparison with numerical results. Based on this, several workspace properties are shown for point-mass cable robots. Finally, it is shown how this workspace-generation approach can be used to analytically formulate other workspaces

Cable-driven robots and multifinger grasps are closely related to each other by the fact that both systems can provide unidirectional forces to the object to be constrained. While prior research primarily related cable-driven robots to grasps without friction, this article seeks to establish relationships between certain cable-driven robots and grasps with friction. Those relationships are used to transfer the antipodal grasp theorem to cable-driven robots, resulting in the so-called antipodal cable theorem. A planar and spatial version are presented.

Singular configurations where a parallel manipulator can lose stiffness are well documented and studied in literature. There exist many proposed measures on how close a pose is to these singularities. This paper develops a framework which unites many of the existing measures, provides further understanding for others, and creates new ones. The framework originates from a physical understanding of singularities with a natural progression to a constrained optimization problem. The solution to the optimization problem is a simple general eigenvalue problem of constant matrices. The power of this new framework is that it is easy to compare existing measures and develop new ones. Several examples of measures within the framework are given.

This paper introduces a slope-based measure of the stability of a pose of an underconstrained cable robot. In order to define this measure, the set of twists that the end-effector can instantaneously undergo is found. The measure is then derived using an energy-based approach to generalize the definition of slope for mixed-dimensional twists. In order to more easily understand and calculate the measure an intermediate space is introduced. An example is worked out for a sample manipulator and the stability measure is used to define a measure of overall manipulator stability.

This paper introduces the available net wrench set, which is the set of all wrenches that a cable robot can apply to its surroundings without violating tension limits in the cables. This set is used as a framework for understanding and computing wrench-feasibility of a pose of the robot, allowing wrench-feasibility to be tested in the task space. The geometric properties of the available net wrench set are then exploited to permit simple geometric calculation of the boundaries of the wrench-feasible workspace. This workspace generation is extended to two other workspaces as well. Other design tools, including payload specification and failure analysis, are also presented.

Although cable-driven robots and parallel manipulators have very similar architecture, the presence of the unidirectional constraints in cable-driven robots makes it impossible to apply many of the concepts and methods used for parallel manipulators. Instead many tools from grasping are more suitable, since fingers also provide unidirectional constraints. This article reviews some of the connections, provides some basic definitions that we hope leads to standardized terminology for cable-driven robots, and finally transfers the antipodal grasping theorem to planar cable-driven robots.

Cable-actuated manipulators possess a number of unique properties, which can have complex effects on their wrench-exertion and stability characteristics. Consequently, their useful workspaces are often unknown or irregularly shaped, so such robots are typically restricted to operate in unnecessarily small, conservatively estimated workspaces. To describe this "useful" workspace, this paper introduces the force-feasible workspace (FFW), defined as the set of end-effector poses for which static equilibrium is maintained and all of a set of required wrenches can be exerted on the environment. Next, this paper presents a means for analytically deriving this region for the case of a point-mass end-effector, which has been validated via numerical simulation. Characteristics and trends of the FFW are observed and analyzed, and correlations between robot design parameters and observed workspace properties are determined. The analysis presented here is a valuable tool for the design of cable robots with optimized workspace properties.

Singularities of parallel manipulators cause severe problems in the use of such manipulators. Likewise, finding how close the manipulator is to a singularity is key to its operation. This article outlines a new method to define "closeness" to singularities based on constrained optimization and its resulting general eigenvalue problem. This framework joins together many existing techniques into one unified theory. From this framework, the similarities and differences between measures can be easily identified, and new measures (e.g. natural frequency) can be derived.

The goal of this research is to enhance the capabilities of transport vehicles so that they can carry delicate objects without extensive packaging. The proposed approach, as first described by Graf and Dillmann (1997), is to add a robotic device (e.g., a parallel manipulator) on top of the vehicle, the motion of which compensates for certain forces and moments. This approach is called active acceleration compensation. This paper discusses performance measures for the effectiveness of acceleration compensation and a new motion planning scheme, the augmented pendulum algorithm, that is particularly effective. The basic idea is to emulate the motion of a free-swinging pendulum to minimize lateral forces acting on the objects.

The paper presents the spherical joint mechanism, a new mechanism for implementing multiple collocated spherical joints. The mechanism is designed to be strong, rigid, easy to build and have a large range of motion. Two different design concepts are presented and compared to other implementations of spherical joints previously proposed. Finally, implementation issues and potential applications of the spherical joint mechanism are discussed.

This paper examines the kinematic design of digital clay, a new type of computer interface that creates 3-D shapes that a user can interact with. After enumerating the design specifications for the digital clay, a new set of kinematic architectures is introduced: the formable crust architectures. These formable crust mechanisms can form shapes in much the same way a piece of fabric would and have tremendous potential for shape generation. Two measures, the surface freedom measure and the freedom from ground measure, are introduced to aid in evaluating potential designs. Finally, the various architectures are evaluated and compared.

The NSF Digital Clay project is focused on the design, prototyping, integration, and validation of a computer-controlled physical device capable of taking any of a wide range of possible shapes in response to changes in a digital 3D model or to changes in the pressure exercised upon it by human hands. Although it clearly is a natural and unavoidable evolution of 3D graphical user interfaces, its unprecedented capabilities constitute a major leap in technologies and paradigms for 3D display, for 3D input, and for collaborative 3D design. In this paper, we provide an overview of the Digital Clay project and discuss the challenges, design choices, and initial solutions for a new finger sculpting interface designed for the Digital Clay and prototyped using conventional 3D I/O hardware.

At a singular configuration the end-effector of a parallel manipulator may become underconstrained, i.e. instantaneous motion may become possible. The unconstrained motion may be only instantaneous, or it can result in finite motion, also known as self-motion. This article presents two different approaches to analyze the potential for large motion at a singular configuration: (1) A necessary (but not sufficient) condition for self-motion is derived that assumes that the manipulator has perfect precision; (2) The unconstrained motion is analyzed based on finite clearance in the joints. Numerical results are presented and compared.

It is well known that parallel manipulators may encounter singular configurations that can result in a loss of full constraint. The nature of the kinematic deficiency at such a singular configuration can be analyzed by calculating the nullspace of the Jacobian matrix at that configuration. The result is a set of one or more screws, called the EE deficiency twists, representing the unconstrained instantaneous motion of the end-effector that cannot be controlled through the actuators. This article discusses the importance of the EE deficiency twists: 1) to identify singular configurations whose unconstrained motion leads directly into a new singular configuration; and 2) for the synthesis of redundant actuation in order to eliminate singular configurations.

The goal of this research is to enhance the capabilities of transport vehicles so that they can carry delicate objects of various shapes and sizes without requiring extensive packaging to protect them. This will be achieved, as first proposed by Graf and Dillmann (1997), by mounting a robotic device on top of the vehicle whose motion compensates for any forces or torques that act on the objects as a result of the vehicle's motion, including disturbances caused by uneven terrain. This approach is called "active acceleration compensation". Several different approaches for the motion planning are implemented and compared. One degree-of-freedom motion planning algorithms include: (1) an algorithm based on optimal control theory; (2) an algorithm based on global optimization schemes; and (3) a more flexible local optimization scheme based on feedback algorithms (FMPA). Finally, a three degrees-of-freedom motion planning algorithm, based on a combination of FMPA and a so called pendulum algorithm, is presented. Simulation, results show that active acceleration compensation using these methods has the potential to significantly improve the performance of the system, as compared to (a) using no actuation at all; or (b) using the classic wash-out filter for motion planning.

This article discusses the modeling of the dynamics of a class of spatial parallel platform mechanisms. For mechanisms of this class perfect static balancing can be achieved through the use of springs. This means that zero actuator torques are required whenever the mechanism is at rest. Furthermore, only inertial forces have to be sustained when the manipulator is moving. Models for the dynamics are implemented based on two different methods, the principle of virtual work and the method of Neutron-Euler. The results of the dynamic analysis can be used to calculate and compare the required actuator torques of different mechanisms for specific payload and trajectories. The ultimate goal is to use these results to design a prototype with six degrees-of-freedom that only requires very small actuator torques to be used for example in flight simulators.

The concept of a convolution product of real-valued functions on the special Euclidean group is applied to the determination of workspaces of discretely actuated manipulators. If a discretely actuated manipulator has n actuators each with K states, then it can reach K/sup n/ frames in space. Given this exponential growth in the number of reachable frames, brute force representation of discretely actuated manipulator workspaces is not feasible in the highly actuated case. However, by partitioning a discretely actuated manipulator into segments, and approximating the workspace of each segment as a density function on a compact subset of the special Euclidean group, the whole workspace can be approximated as an n-fold convolution of these densities. A numerical approximation of this convolution is presented in this paper that requires O(logn) computation time as compared to the O(K/sup n/) computations required by brute force workspace generation.

In this work, the concept of a convolution product of real-valued functions on the special Euclidean group, SE(D) (which describes all rigid body motions in D-dimensional Euclidean space), is applied to the determination of workspaces of discretely actuated manipulators. These manipulators have a finite number of joint states. If a discretely actuated manipulator consists of P actuated modules, each with K states, then it can reach K/sup p/ frames in space. Given this exponential growth in the number of reachable frames, brute force representation of discretely actuated manipulator workspaces is not feasible in the highly actuated case. However, by partitioning a discretely actuated manipulator into P modules, and approximating the workspace of each module as a density function on a compact subset of the special Euclidean group, the whole workspace can be approximated as an P-fold convolution of these densities. A numerical approximation of this convolution is presented in this paper which is O(P) for fixed taskspace dimension. In the special case when the manipulator is composed of P identical actuated modules, the workspace density for the whole manipulator can be calculated in O(log P) computation time. In either case, the O(K/sup p/) computations required by brute force workspace generation are avoided.

In this paper the problem of dynamic self-reconfiguration of a class of modular robotic systems referred to as metamorphic systems is examined. A metamorphic robotic system is a collection of mechatronic modules, each of which has the ability to connect, disconnect, and climb over adjacent modules. We examine the near-optimal reconfiguration of a metamorphic robot from an arbitrary initial configuration to a desired final configuration. Concepts of distance between metamorphic robot configurations are defined, and shown to satisfy the formal properties of a metric. These metrics, called configuration metrics, are then applied to the automatic self-reconfiguration of metamorphic systems in the case when one module is allowed to move at a time. There is no simple method for computing the optimal sequence of moves required to reconfigure. As a result, heuristics which can give a near optimal solution must be used. We use the technique of simulated annealing to drive the reconfiguration process with configuration metrics as cost functions. The relative performance of simulated annealing with different cost functions is compared and the usefulness of the metrics developed in this paper is demonstrated.

Hyper-redundant manipulators present an alternative to conventional 6 DOF manipulators for inspection, space, and medical applications. The additional degrees of freedom facilitate obstacle avoidance and allow tasks to be performed even if some of the actuators fail. In this paper the authors consider hyper-redundant manipulators that are actuated discretely, e.g. using two-state actuators or motors with finite resolution. The inverse kinematics problem for a discretely actuated manipulator is intrinsically different from the one for its continuously actuated counterpart. The authors present a framework for the discussion of the discretely actuated case and propose an algorithm for the inverse kinematics. The algorithm generates solutions in linear time with respect to the number of manipulator actuators, as opposed to the exponential time required by brute force search.