We describe our full body humanoid control approach developed for the simulation phase of the DARPA Robotics Challenge (DRC), as well as the modifications made for the DARPA Robotics Challenge Trials. We worked with the Boston Dynamics Atlas robot. Our approach was initially targeted at walking, and it consisted of two levels of optimization: a high-level trajectory optimizer that reasons about center of mass and swing foot trajectories, and a low-level controller that tracks those trajectories by solving floating base full body inverse dynamics using quadratic programming. This controller is capable of walking on rough terrain, and it also achieves long footsteps, fast walking speeds, and heel-strike and toe-off in simulation. During development of these and other whole body tasks on the physical robot, we introduced an additional optimization component in the low-level controller, namely an inverse kinematics controller. Modeling and torque measurement errors and hardware features of the Atlas robot led us to this three-part approach, which was applied to three tasks in the DRC Trials in December 2013. C 2014 Wiley Periodicals, Inc.
One popular approach to controlling humanoid robots is through inverse kinematics (IK) with stiff joint position tracking. On the other hand, inverse dynamics (ID) based approaches have gained increasing acceptance by providing compliant motions and robustness to external perturbations. However, the performance of such methods is heavily dependent on high quality dynamic models, which are often very difficult to produce for a physical robot. IK approaches only require kinematic models, which are much easier to generate in practice. In this paper, we supplement our previous work with ID-based controllers by adding IK, which helps compensate for modeling errors. The proposed full body controller is applied to three tasks in the DARPA Robotics Challenge (DRC) Trials in Dec. 2013.
Background and objectiveMost previous studies adopted single traditional time series models to predict incidences of malaria. A single model cannot effectively capture all the properties of the data structure. However, a stacking architecture can solve this problem by combining distinct algorithms and models. This study compares the performance of traditional time series models and deep learning algorithms in malaria case prediction and explores the application value of stacking methods in the field of infectious disease prediction.MethodsThe ARIMA, STL+ARIMA, BP-ANN and LSTM network models were separately applied in simulations using malaria data and meteorological data in Yunnan Province from 2011 to 2017. We compared the predictive performance of each model through evaluation measures: RMSE, MASE, MAD. In addition, gradient-boosting regression trees (GBRTs) were used to combine the above four models. We also determined whether stacking structure improved the model prediction performance.ResultsThe root mean square errors (RMSEs) of the four sub-models were 13.176, 14.543, 9.571 and 7.208; the mean absolute scaled errors (MASEs) were 0.469, 0.472, 0.296 and 0.266 and the mean absolute deviation (MAD) were 6.403, 7.658, 5.871 and 5.691. After using the stacking architecture combined with the above four models, the RMSE, MASE and MAD values of the ensemble model decreased to 6.810, 0.224 and 4.625, respectively.ConclusionsA novel ensemble model based on the robustness of structured prediction and model combination through stacking was developed. The findings suggest that the predictive performance of the final model is superior to that of the other four sub-models, indicating that stacking architecture may have significant implications in infectious disease prediction.
This study addresses the problem of learning robust frame-level feature representation for unsupervised subword modeling in the zero-resource scenario. Robustness of the learned features is achieved through effective speaker adaptation and exploiting cross-lingual phonetic knowledge. For speaker adaptation, an out-of-domain automatic speech recognition (ASR) system is used to estimate fMLLR features for untranscribed speech of target zero-resource languages. The fMLLR features are applied in multi-task learning of a deep neural network (DNN) to further obtain phonetically discriminative and speaker-invariant bottleneck features (BNFs). Frame-level labels for DNN training can be acquired based on two approaches: Dirichlet process Gaussian mixture model (DPGMM) clustering, and outof-domain ASR decoding. Moreover, system fusion is performed by concatenating BNFs extracted by different DNNs. Our methods are evaluated by ZeroSpeech 2017 Track one, where the performance is evaluated by ABX minimal pair discriminability. Experimental results demonstrate that: (1) Using an out-of-domain ASR system to perform speaker adaptation of zero-resource speech is effective and efficient; (2) Our system achieves highly competitive performance to state of the art; (3) System fusion could improve feature representation capability.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.