This paper addresses a vision-based 3D motion estimation framework for humanoid robots, which copes with human-like walking pattern. A humanoid robot, called Roboray, is designed for dynamic walking control with heel-toe motion like a human. In spite of stability and energy efficiency of the dynamic walking, it accompanies larger swaying motion and more uncertainty in camera movement than the conventional ZMP (Zero Moment Point)-based walking does. The framework effectively uses on-board odometry information from the robot to improve the performance of the visionbased motion estimation. To accomplish this, we propose an onboard odometry filter which fuses kinematic odometry, visual odometry, and raw IMU data. And the odometry filter is combined with vision-based SLAM to provide accurate motion model, so it enhances the SLAM estimates. Experimental results in indoor environment verify that the framework can successfully estimate the pose of Roboray in real-time.

This paper deals with the practical problem as to how a humanoid robot opens a door in an office environment. Among the several tasks for the door opening, we focus on the recognition of a door and grasping its knob and propose a systematic touch scheme that combines the visual and touch data to grasp a knob safely and robustly. The visual recognition of a door and its knob is done by feature matching by a feature classifier and a segmentation method with minor assumptions about a knob, respectively. The position of a knob from the visual recognition is refined by the proposed touch scheme where a robot touches a knob several times with a hand equipped with a force-torque sensor. We realize that in the on-line experiment, a robot recognizes a door and its knob, and grasps the knob and partially opens the door.

This paper deals with the practical problem of how a humanoid robot recognizes a door and its knob in an office environment. This is the initial work where a humanoid robot recognizes, approaches, and opens a door. We propose an integrated solution for visual recognition of a door and its knob with minor constraints. In our approach, the door recognition is transformed into the classification problem of feature points, which can be done quickly enough to conduct it on-line. The knob is extracted by a segmentation method and a few thresholds: blob size, blob ratio, and the distance from the floor. We show in the experiment that our humanoid robot can recognize a door and its knob reliably and quickly.

FastSLAM has been shown to degenerate over time due to sample impoverishment, that is, poor samples are generated during the sampling process. One of major culprits of the sample impoverishment problem is lack of the number of particles estimating the pose of the robot and the environment. In this work, an adaptive prior boosting technique is proposed for the efficient sample size according to the uncertainty of each situation in performing FastSLAM. It uses a back-propagation neural network, learned in various environments, in order to decide the required sample size. This adaptive approach generates a small number of particles when the uncertainty is low while performing FastSLAM, and it generates a large number of particles when the uncertainty is high. This technique efficiently generates the sample size in computer simulations and real environmental experiments, which significantly reduces the RMS feature and position errors.

The state-of-the-art FastSLAM has been shown to cause a particle depletion problem while performing simultaneous localization and mapping for mobile robots. As a result, it always produces over-confident estimates of uncertainty as time progresses. This particle depletion problem is mainly due to the resampling process in FastSLAM, which tends to eliminate particles with low weights. Therefore, the number of particles to perform loop-closure decreases, which makes the performance of FastSLAM degenerate. The resampling process has not been thoroughly analyzed even though it is the main reason for the particle depletion problem. In this paper, standard resampling algorithms (systematic residual and partial resampling), and a rank-based resampling applying genetic algorithms are analyzed using computer simulations. For the thorough analysis, several performance measures are used such as effective sample size, the number of distinct particles, root mean square (RMS) errors, and complexity. According to the simulation results, all resampling algorithms could not resolve the particle depletion problem.

We propose a low cost, accurate and safe active range sensor with a halogen sheet-of-light and a USB camera. We also propose an algorithm suitable for building a local map using the probabilistic sensor model. In this paper, the inexpensive structured light source that uses tungsten-halogen lamp is used, and an efficient method of detecting the sheet-of-light by the camera is proposed. A local map is built using the spatial occupancy technique which is inputs values in an occupancy grid that reflect the probability of occupancy for each cell. In addition, a suitable probabilistic sensor model is employed to build the local map. The experimental results show the validity of the proposed range sensor and the algorithm to build a map using the probabilistic sensor model.