In this study, we present a method to grasp diverse unseen real-world objects using an off-policy actor-critic deep reinforcement learning (RL) with the help of a simulation and the use of as little real-world data as possible. Actor-critic deep RL is unstable and difficult to tune when a raw image is given as an input. Therefore, we use state representation learning (SRL) to make actor-critic RL feasible for visual grasping tasks. Meanwhile, to reduce visual reality gap between simulation and reality, we also employ a typical pixel-level domain adaptation that can map simulated images to realistic ones. In our method, as the SRL model is a common preprocessing module for simulated and real-world data, we perform SRL using real and adapted images. This pixel-level domain adaptation enables the robot to learn grasping skills in a real environment using small amounts of real-world data. However, the controller trained in the simulation should adapt to the real world efficiently. Hence, we propose a method combining a typical pixel-level domain adaptation and the proposed SRL model, where we perform SRL based on a feature-level domain adaptation. In evaluations of vision-based robotics grasping tasks, we show that the proposed method achieves a substantial improvement over a method that only employs a pixel-level or domain adaptation.

For robotic grasping tasks with diverse target objects, some deep learning-based methods have achieved state-of-the-art results using direct visual input. In contrast, actor-critic deep reinforcement learning (RL) methods typically perform very poorly when applied to grasp diverse objects, especially when learning from raw images and sparse rewards. To render these RL techniques feasible for vision-based grasping tasks, we used state representation learning (SRL), in which we encode essential information for subsequent use in RL. However, typical representation learning procedures are unsuitable for extracting pertinent information for learning grasping skills owing to the high complexity of visual inputs for representation learning, in which a robot attempts to grasp a target object. We found that the proposed preprocessed input image is the key to capturing effectively a compact representation. This enables deep RL to learn robotic grasping skills from highly varied and diverse visual inputs. Further, we demonstrate the effectiveness of the proposed approach with varying levels of preprocessing in a realistic simulated environment. We also describe how the resulting model can be transferred to a real-world robot and also demonstrate a 68% success rate on real-world grasp attempts.

In this study, we aim at grasping a single target object in a cluttered environment using a robotic arm. While dexterous grasp for various shapes of objects is not considered in this work, we focus on developing the method to mitigate clutter near the target object as soon as quickly. For this purpose, we propose a method to generate nonlinear pushing motions for object singulation based on an off-the-shelf machine learning algorithm and a typical semantic segmentation algorithm. Through experiments, we show that the success rate of robotic grasping is considerably improved by the proposed pushing behavior. And notably, the nonlinear pushing trajectories allows the robot to perform singulation of the target object in a cluttered environment with fewer trials than linear pushing usually pursued in related works.

Object recognition is one of the most important applications of robotics. For object recognition, edge orientation is widely used as a primitive visual feature. However, a classical filter-based approach passes not only edges inside target orientation band but also edges outside. This can thus cause problem in the estimation of the true orientation of edge. This study proposes a filtering scheme to reduce the false-positive responses, i.e. edges outside target orientation band, and investigate a solution inspired by biological vision. Motivated by several psychophysical and neuro-physiological findings, we present a computational framework based on the basic mechanisms of cortical processing, i.e. feed-forward, lateral and feedback stages. In the feed-forward stage, our model uses a classical filter-based method to allow as many true orientation edges to pass through as possible. False responses in orientation channels are then inhibited by lateral interaction. The remaining undesired responses are suppressed through the feedback stage. We evaluated the performance of our model against classical filter-based methods such as Gabor and Neumann filtering using several artificial and natural images. The results validated the effectiveness of our approach.

Disparity computation in occluded or texture-less regions is considered to be a fundamental issue in dense stereo matching, but there is another practical issue that must be resolved before it can be used effectively in various robotics applications. This issue is the problem of intensity difference between corresponding pixels of an image pair. To tackle such problems, we present a dependable stereo matching algorithm using two-layer recurrent process and chaining search. Two-layer process integrates pixel and region-levels information through recurrent interaction. To estimate the precise disparities in occluded regions, reliable disparities in non-occluded region are propagated to occluded regions by the proposed chaining search. To test our algorithm, it was compared with two outstanding algorithms in Middlebury benchmark using Gaussian noisy images. The results validated the effectiveness of our approach.

In this paper, to provide a robot with information relative to structure of its environment, we propose a method to recognize types of structural corridor landmarks such as T-junction, L-junction, end of the corridor, using vanishing points-based visual image features and hidden Markov models. Several experimental results are illustrated to demonstrate the validity of the proposed approach in a real environment.

This paper presents a predictive visual recognition method to identify types of structural corridor landmark such as T-junctions, L-junctions and ends of the corridors, using vanishing point-based visual features and a two-layer HMM framework. The main advantage of the approach is that a mobile robot can recognize the types of structural landmark before reaching them. This is important for efficient robot navigation because it allows the robot to make navigation decision in advance. Several experimental results are provided to demonstrate the validity of the proposed approach in a real environment.

This work considers robot localization with an action-associated sparse appearance-based map, under conditions with dynamic change in the environment. In this case, two significant problems must be solved for robust localization. The first involves variations in the environment caused by dynamic objects and changes in illumination, and the second arises from the nature of sparse appearance-based map. That is, a robot must be able to recognize observations taken at slightly different positions and angles within a certain region as identical. In this paper, we address a possible solution to these problems on the basis of a probabilistic model called the Bayes filter. Here, we propose an observation model based LeTO2function and an action-associated sparse appearance-based map to be used for prediction, update, and final localization steps. In addition, multiple visual features are used to increase the reliability of the observation model. We performed experiments to demonstrate the validity of the proposed approach under various conditions with regard to dynamic objects, illumination, and viewpoint. The results clearly demonstrated the value of our approach.

Task planning by hierarchical abstraction of world elements and behaviors(PhaWeB) is proposed for mobile robots task planning, where abstraction of world elements is used to reduce search space and abstraction of behaviors is used to efficiently find more detailed behaviors. To integrate abstraction of world elements with abstraction of behaviors, we construct hierarchical domain description based on joint of action hierarchy (JAH). In JAH, some actions in a level of action hierarchy are included in the adjacent level of action hierarchy. To show the validities of our proposed PhaWeB, experimental results are illustrated and will be compared with world elements abstraction based planning approach as well as behaviors abstraction based planning approach.