-
Distributed Reinforcement Learning of Targeted Gras** with Active Vision for Mobile Manipulators
Authors:
Yasuhiro Fujita,
Kota Uenishi,
Avinash Ummadisingu,
Prabhat Nagarajan,
Shimpei Masuda,
Mario Ynocente Castro
Abstract:
Develo** personal robots that can perform a diverse range of manipulation tasks in unstructured environments necessitates solving several challenges for robotic gras** systems. We take a step towards this broader goal by presenting the first RL-based system, to our knowledge, for a mobile manipulator that can (a) achieve targeted gras** generalizing to unseen target objects, (b) learn comple…
▽ More
Develo** personal robots that can perform a diverse range of manipulation tasks in unstructured environments necessitates solving several challenges for robotic gras** systems. We take a step towards this broader goal by presenting the first RL-based system, to our knowledge, for a mobile manipulator that can (a) achieve targeted gras** generalizing to unseen target objects, (b) learn complex gras** strategies for cluttered scenes with occluded objects, and (c) perform active vision through its movable wrist camera to better locate objects. The system is informed of the desired target object in the form of a single, arbitrary-pose RGB image of that object, enabling the system to generalize to unseen objects without retraining. To achieve such a system, we combine several advances in deep reinforcement learning and present a large-scale distributed training system using synchronous SGD that seamlessly scales to multi-node, multi-GPU infrastructure to make rapid prototy** easier. We train and evaluate our system in a simulated environment, identify key components for improving performance, analyze its behaviors, and transfer to a real-world setup.
△ Less
Submitted 14 October, 2020; v1 submitted 15 July, 2020;
originally announced July 2020.
-
Chainer: A Deep Learning Framework for Accelerating the Research Cycle
Authors:
Seiya Tokui,
Ryosuke Okuta,
Takuya Akiba,
Yusuke Niitani,
Toru Ogawa,
Shunta Saito,
Shuji Suzuki,
Kota Uenishi,
Brian Vogel,
Hiroyuki Yamazaki Vincent
Abstract:
Software frameworks for neural networks play a key role in the development and application of deep learning methods. In this paper, we introduce the Chainer framework, which intends to provide a flexible, intuitive, and high performance means of implementing the full range of deep learning models needed by researchers and practitioners. Chainer provides acceleration using Graphics Processing Units…
▽ More
Software frameworks for neural networks play a key role in the development and application of deep learning methods. In this paper, we introduce the Chainer framework, which intends to provide a flexible, intuitive, and high performance means of implementing the full range of deep learning models needed by researchers and practitioners. Chainer provides acceleration using Graphics Processing Units with a familiar NumPy-like API through CuPy, supports general and dynamic models in Python through Define-by-Run, and also provides add-on packages for state-of-the-art computer vision models as well as distributed training.
△ Less
Submitted 1 August, 2019;
originally announced August 2019.
-
Evidence for a New Magnetoelectric Effect of Current-Induced Magnetization in a Toroidal Magnetic Ordered State of UNi$_4$B
Authors:
Hiraku Saito,
Kenta Uenishi,
Naoyuki Miura,
Chihiro Tabata,
Hiroyuki Hidaka,
Tatsuya Yanagisawa,
Hiroshi Amitsuka
Abstract:
Magnetization measurements under direct electric currents were performed for toroidal magnetic ordered state of UNi$_4$B to test a recent theoretical prediction of current-induced magnetization in a metallic system lacking local inversion symmetry.We found that each of the electric currents parallel to [$2\bar{1}\bar{1}0$] and [$0001$] in the hexagonal 4-index notation induces uniform magnetizatio…
▽ More
Magnetization measurements under direct electric currents were performed for toroidal magnetic ordered state of UNi$_4$B to test a recent theoretical prediction of current-induced magnetization in a metallic system lacking local inversion symmetry.We found that each of the electric currents parallel to [$2\bar{1}\bar{1}0$] and [$0001$] in the hexagonal 4-index notation induces uniform magnetization in the direction of [$01\bar{1}0$].The observed behavior of the induced magnetization is essentially consistent with the theoretical prediction; however it also shows an inconsistency suggesting that the antiferromagnetic state of UNi$_4$B could not simply be regarded as a uniform toroidal order in the ideal honeycomb layered structure.
△ Less
Submitted 14 January, 2018;
originally announced January 2018.
-
Synthesis and physical properties of a new caged compound Ce3Pd20As6 of the C6Cr23-type structure
Authors:
K. Orita,
K. Uenishi,
M. Tsubota,
Y. Shimada,
T. Onimaru,
T. Takabatake,
J. Kitagawa
Abstract:
We have found that Ce3Pd20As6 crystallizes into a cubic C6Cr23-type structure. Combination of electron probe microanalysis of the chemical composition and Rietveld analysis of the powder X-ray diffraction pattern has revealed an inhomogeneous atomic composition of variable stoichiometry. The physical properties of Ce3Pd20As6 were investigated by measuring the magnetization, electrical resistivity…
▽ More
We have found that Ce3Pd20As6 crystallizes into a cubic C6Cr23-type structure. Combination of electron probe microanalysis of the chemical composition and Rietveld analysis of the powder X-ray diffraction pattern has revealed an inhomogeneous atomic composition of variable stoichiometry. The physical properties of Ce3Pd20As6 were investigated by measuring the magnetization, electrical resistivity and specific heat. The 4f electrons of Ce3+ ions are well localized but do not show phase transition down to 0.5 K. The metallic electrical resistivity shows a weak Kondo screening. The residual resistivity ratio is rather low probably due to the variable stoichiometry. The magnetization curve and magnetic entropy suggest the Gamma_8 quartet crystalline-electric-field ground state at least one of two Ce sites.
△ Less
Submitted 28 October, 2014;
originally announced October 2014.
