Abstract Vision-based haptic sensors have emerged as a promising approach to robotic touch due affordable high-resolution cameras and successful computer vision techniques; however, their physical design the information they provide do not yet meet requirements of real applications. We present robust, soft, low-cost, vision-based, thumb-sized three-dimensional sensor named Insight, which continually provides directional force-distribution map over its entire conical sensing surface....

Physics is a field of science that has traditionally used the scientific method to answer questions about why natural phenomena occur and make testable models explain phenomena. Discovering equations, laws, principles are invariant, robust, causal been fundamental in physical sciences throughout centuries. Discoveries emerge from observing world and, when possible, performing interventions on system under study. With advent big data data-driven methods, fields equation discovery have...

The Variational Autoencoder (VAE) is a powerful architecture capable of representation learning and generative modeling. When it comes to interpretable (disentangled) representations, VAE its variants show unparalleled performance. However, the reasons for this are unclear, since very particular alignment latent embedding needed but design does not encourage in any explicit way. We address matter offer following explanation: diagonal approximation encoder together with inherent stochasticity...

Achieving fusion of deep learning with combinatorial algorithms promises transformative changes to artificial intelligence. One possible approach is introduce building blocks into neural networks. Such end-to-end architectures have the potential tackle problems on raw input data such as ensuring global consistency in multi-object tracking or route planning maps robotics. In this work, we present a method that implements an efficient backward pass through blackbox implementations solvers...

Anisotropic patchy particles have become an archetypical statistical model system for associating fluids. Here, we formulate approach to the Kern–Frenkel via classical density functional theory describe positionally and orientationally resolved equilibrium distributions in flat wall geometries. The is split into a reference part averaged orientational mean-field approximation. To bring kernel form suitable machine learning (ML) techniques, expansion invariants proper incorporation of...

Information theory is a powerful tool to express principles drive autonomous systems because it domain invariant and allows for an intuitive interpretation. This paper studies the use of predictive information (PI), also called excess entropy or effective measure complexity, sensorimotor process as driving force generate behavior. We study nonlinear nonstationary introduce time-local predicting (TiPI) which us derive exact results together with explicit update rules parameters controller in...

Event-triggered control (ETC) methods can achieve high-performance with a significantly lower number of samples compared to usual, time-triggered methods. These frameworks are often based on mathematical model the system and specific designs controller event trigger. In this paper, we show how deep reinforcement learning (DRL) algorithms be leveraged simultaneously learn communication behavior from scratch, present DRL approach that is particularly suitable for ETC. To our knowledge, first...

Rank-based metrics are some of the most widely used criteria for performance evaluation computer vision models. Despite years effort, direct optimization these remains a challenge due to their non-differentiable and non-decomposable nature. We present an efficient, theoretically sound, general method differentiating rank-based with mini-batch gradient descent. In addition, we address instability sparsity supervision signal that both arise from using as targets. Resulting losses based on...

Tactile feedback is essential to make robots more agile and effective in unstructured environments. However, high-resolution tactile skins are not widely available; this due the large size of robust sensing units because many typically lead fragility wiring high costs. One route toward involves embedding a few sensor (taxels) into flexible surface material use signal processing achieve with superresolution accuracy. Here, we propose theory for geometric guide development sensors kind link it...

Electrical resistance tomography (ERT) can be used to create large-scale soft tactile sensors that are flexible and robust. Good performance requires a fast accurate mapping from the sensor's sequential voltage measurements distribution of force across its surface. However, particularly with multiple contacts, this task is challenging for both previously developed approaches: physics-based modeling end-to-end data-driven learning. Some promising results were recently achieved using...

Intelligent interaction with the physical world requires perceptual abilities beyond vision and hearing; vibrant tactile sensing is essential for autonomous robots to dexterously manipulate unfamiliar objects or safely contact humans. Therefore, robotic manipulators need high‐resolution touch sensors that are compact, robust, inexpensive, efficient. The soft vision‐based haptic sensor presented herein a miniaturized optimized version of previously published Insight. Minsight has size shape...

We present an approach to identify concise equations from data using a shallow neural network approach. In contrast ordinary black-box regression, this allows understanding functional relations and generalizing them observed unseen parts of the parameter space. show how extend class learnable for recently proposed equation learning include divisions, we improve model selection strategy be useful challenging real-world data. For systems governed by analytical expressions, our method can in...

Retina is a paradigmatic system for studying sensory encoding: the transformation of light into spiking activity ganglion cells. The inverse problem, where stimulus reconstructed from spikes, has received less attention, especially complex stimuli that should be "pixel-by-pixel". We recorded around hundred neurons dense patch in rat retina and decoded movies multiple small randomly-moving discs. constructed nonlinear (kernelized neural network) decoders improved significantly over linear...

In classical machine learning, regression is treated as a black box process of identifying suitable function from hypothesis set without attempting to gain insight into the mechanism connecting inputs and outputs. natural sciences, however, finding an interpretable for phenomenon prime goal it allows understand generalize results. This paper proposes novel type learning network, called equation learner (EQL), that can learn analytical expressions able extrapolate unseen domains. It...

Capturing aleatoric uncertainty is a critical part of many machine learning systems. In deep learning, common approach to this end train neural network estimate the parameters heteroscedastic Gaussian distribution by maximizing logarithm likelihood function under observed data. work, we examine and identify potential hazards associated with use log-likelihood in conjunction gradient-based optimizers. First, present synthetic example illustrating how can lead very poor but stable parameter...

While many algorithms for diversity maximization under imitation constraints are online in nature, applications require offline without environment interactions. Tackling this problem the setting, however, presents significant challenges that non-trivial, multi-stage optimization processes with non-stationary rewards. In work, we present a novel algorithm enhances using an objective based on Van der Waals (VdW) force and successor features, eliminates need to learn previously used skill...

We explore the feasibility of using machine learning methods to obtain an analytic form classical free energy functional for two model fluids, hard rods and Lennard-Jones, in one dimension. The equation network proposed by Martius Lampert [e-print arXiv:1610.02995 (2016)] is suitably modified construct densities which are functions a set weighted built from small number basis with flexible combination rules. This setup considerably enlarges space used optimization as compared previous work...

Learning diverse skills is one of the main challenges in robotics. To this end, imitation learning approaches have achieved impressive results. These methods require explicitly labeled datasets or assume consistent skill execution to enable and active control individual behaviors, which limits their applicability. In work, we propose a cooperative adversarial method for obtaining single versatile policies with controllable sets from unlabeled containing state transition patterns by...

Reconstructing the effective equation of motion for time evolution a subset degrees freedom larger system remains problem interest in quantum physics. Many methods have been developed, but they either rely on an ad hoc ansatz, demand data that is not experimentally accessible, or lack physical interpretability. The authors employ machine-learning to infer dynamical generator from noisy, finite set local measurements. Their method yields interpretable results may be used noise models...

Deep learning has recently gained immense popularity in the Earth sciences as it enables us to formulate purely data-driven models of complex system processes. learning-based weather prediction (DLWP) have made significant progress last few years, achieving forecast skills comparable established numerical (NWP) with comparatively lesser computational costs. In order train accurate, reliable, and tractable DLWP several millions parameters, model design needs incorporate suitable inductive...