Imitation Learning (IL) is an effective framework to learn visuomotor skills from offline demonstration data. However, IL methods often fail to generalize to new scene configurations not covered by training data. On the other hand, humans can manipulate objects in varying conditions. Key to such capability is hand-eye coordination, a cognitive ability that enables humans to adaptively direct their movements at task-relevant objects and be invariant to the objects' absolute spatial location. In this work, we present a learnable action space, Hand-eye Action Networks (HAN), that can approximate human's hand-eye coordination behaviors by learning from human teleoperated demonstrations. Through a set of challenging multi-stage manipulation tasks, we show that a visuomotor policy equipped with HAN is able to inherit the key spatial invariance property of hand-eye coordination and achieve zero-shot generalization to new scene configurations. Additional materials available at https://sites.google.com/stanford.edu/han
Learning an empirically effective model with generalization using limited data is a challenging task for deep neural networks. In this paper, we propose a novel learning framework called PurifiedLearning to exploit task-irrelevant features extracted from task-irrelevant labels when training models on small-scale datasets. Particularly, we purify feature representations by using the expression of task-irrelevant information, thus facilitating the learning process of classification. Our work is built on solid theoretical analysis and extensive experiments, which demonstrate the effectiveness of PurifiedLearning. According to the theory we proved, PurifiedLearning is model-agnostic and doesn't have any restrictions on the model needed, so it can be combined with any existing deep neural networks with ease to achieve better performance. The source code of this paper will be available in the future for reproducibility.
The LIght Detection And Ranging (LiDAR) sensor has become one of the most important perceptual devices due to its important role in simultaneous localization and mapping (SLAM). Existing SLAM methods are mainly developed for mechanical LiDAR sensors, which are often adopted by large scale robots. Recently, the solid-state LiDAR is introduced and becomes popular since it provides a cost-effective and lightweight solution for small scale robots. Compared to mechanical LiDAR, solid-state LiDAR sensors have higher update frequency and angular resolution, but also have smaller field of view (FoV), which is very challenging for existing LiDAR SLAM algorithms. Therefore, it is necessary to have a more robust and computationally efficient SLAM method for this new sensing device. To this end, we propose a new SLAM framework for solid-state LiDAR sensors, which involves feature extraction, odometry estimation, and probability map building. The proposed method is evaluated on a warehouse robot and a hand-held device. In the experiments, we demonstrate both the accuracy and efficiency of our method using an Intel L515 solid-state LiDAR. The results show that our method is able to provide precise localization and high quality mapping. We made the source codes public at \url{https://github.com/wh200720041/SSL_SLAM}.
Simultaneous Localization And Mapping (SLAM) is a task to estimate the robot location and to reconstruct the environment based on observation from sensors such as LIght Detection And Ranging (LiDAR) and camera. It is widely used in robotic applications such as autonomous driving and drone delivery. Traditional LiDAR-based SLAM algorithms mainly leverage the geometric features from the scene context, while the intensity information from LiDAR is ignored. Some recent deep-learning-based SLAM algorithms consider intensity features and train the pose estimation network in an end-to-end manner. However, they require significant data collection effort and their generalizability to environments other than the trained one remains unclear. In this paper we introduce intensity features to a SLAM system. And we propose a novel full SLAM framework that leverages both geometry and intensity features. The proposed SLAM involves both intensity-based front-end odometry estimation and intensity-based back-end optimization. Thorough experiments are performed including both outdoor autonomous driving and indoor warehouse robot manipulation. The results show that the proposed method outperforms existing geometric-only LiDAR SLAM methods.
Despite their remarkable expressibility, convolution neural networks (CNNs) still fall short of delivering satisfactory results on single image dehazing, especially in terms of faithful recovery of fine texture details. In this paper, we argue that the inadequacy of conventional CNN-based dehazing methods can be attributed to the fact that the domain of hazy images is too far away from that of clear images, rendering it difficult to train a CNN for learning direct domain shift through an end-to-end manner and recovering texture details simultaneously. To address this issue, we propose to add explicit constraints inside a deep CNN model to guide the restoration process. In contrast to direct learning, the proposed mechanism shifts and narrows the candidate region for the estimation output via multiple confident neighborhoods. Therefore, it is capable of consolidating the expressibility of different architectures, resulting in a more accurate indirect domain shift (IDS) from the hazy images to that of clear images. We also propose two different training schemes, including hard IDS and soft IDS, which further reveal the effectiveness of the proposed method. Our extensive experimental results indicate that the dehazing method based on this mechanism outperforms the state-of-the-arts.
Convolutional Neural Networks(CNNs) has achieved remarkable performance breakthrough in Euclidean structure data. Recently, aggregation-transformation based Graph Neural networks(GNNs) gradually produce a powerful performance on non-Euclidean data. In this paper, we propose a cross-correlation based graph convolution method allowing to naturally generalize CNNs to non-Euclidean domains and inherit the excellent natures of CNNs, such as local filters, parameter sharing, flexible receptive field, etc. Meanwhile, it leverages dynamically generated convolution kernel and cross-correlation operators to address the shortcomings of prior methods based on aggregation-transformation or their approximations. Our method has achieved or matched popular state-of-the-art results across three established graph benchmarks: the Cora, Citeseer, and Pubmed citation network datasets.
Several robot manipulation tasks are extremely sensitive to variations of the physical properties of the manipulated objects. One such task is manipulating objects by using gravity or arm accelerations, increasing the importance of mass, center of mass, and friction information. We present SwingBot, a robot that is able to learn the physical features of a held object through tactile exploration. Two exploration actions (tilting and shaking) provide the tactile information used to create a physical feature embedding space. With this embedding, SwingBot is able to predict the swing angle achieved by a robot performing dynamic swing-up manipulations on a previously unseen object. Using these predictions, it is able to search for the optimal control parameters for a desired swing-up angle. We show that with the learned physical features our end-to-end self-supervised learning pipeline is able to substantially improve the accuracy of swinging up unseen objects. We also show that objects with similar dynamics are closer to each other on the embedding space and that the embedding can be disentangled into values of specific physical properties.
Cancer is responsible for millions of deaths worldwide every year. Although significant progress hasbeen achieved in cancer medicine, many issues remain to be addressed for improving cancer therapy.Appropriate cancer patient stratification is the prerequisite for selecting appropriate treatment plan, ascancer patients are of known heterogeneous genetic make-ups and phenotypic differences. In thisstudy, built upon deep phenotypic characterizations extractable from Mayo Clinic electronic healthrecords (EHRs) and genetic test reports for a collection of cancer patients, we evaluated variousgraph neural networks (GNNs) leveraging a joint of phenotypic and genetic features for cancer typeclassification. Models were applied and fine-tuned on the Mayo Clinic cancer disease dataset. Theassessment was done through the reported accuracy, precision, recall, and F1 values as well as throughF1 scores based on the disease class. Per our evaluation results, GNNs on average outperformed thebaseline models with mean statistics always being higher that those of the baseline models (0.849 vs0.772 for accuracy, 0.858 vs 0.794 for precision, 0.843 vs 0.759 for recall, and 0.843 vs 0.855 for F1score). Among GNNs, ChebNet, GraphSAGE, and TAGCN showed the best performance, while GATshowed the worst. We applied and compared eight GNN models including AGNN, ChebNet, GAT,GCN, GIN, GraphSAGE, SGC, and TAGCN on the Mayo Clinic cancer disease dataset and assessedtheir performance as well as compared them with each other and with more conventional machinelearning models such as decision tree, gradient boosting, multi-layer perceptron, naive bayes, andrandom forest which we used as the baselines.