International Journal of Humanoid Robotics

This paper proposes a self-supervised model which enables a humanoid robot to learn to reach to visual targets. Only 400 training samples are used to learn a forward kinematic model of the six degree-of-freedom (DOF) arm. The forward model is represented compactly with just 150 hidden neurons and enables high accuracy reaching in real time. We provide an optimization process for the learning parameters and a careful analysis of reaching errors. An extension of the model is presented to address additional DOFs in the neck. The consistency of the model with physiological and psychological observations is elaborated.

This paper proposes a learning method for detecting object affordances through haptic exploration by a multi-fingered robot hand. Learning how to remove a screw cap from a bottle is the present target task. Assuming that coarse manipulation strategy is given by other methods, such as visual observation of a model task, the system applies coarse grabbing actions to the target object. In response to the exploratory actions, the object moves (rotates, in this case) along the physical constraint (screw). The robot detects the resulting motion through proprioception of the compliant fingers. A non-supervised statistical learning method is applied to categorize the resulting motion. The method is a combination of high-order local autocorrelation (HLAC), principal components analysis (PCA), and mean-shift clustering. Experiments with a real multi-fingered robot hand and bottle caps of different diameters confirm that the proposed method can detect and categorize rotational constraints.

This paper addresses a construction method of a system that realizes whole body reaching motion of humanoids. Humanoids have many redundant degrees of freedom for reaching, and even the base can be moved by making the robot step. Therefore, there are infinite final posture solutions for a final goal position of reaching, and there are also infinite solutions for reaching trajectories that realize a final reaching posture. It is, however, difficult to find an appropriate solution because of the constraint of dynamic balance, and relatively narrow movable range for each joint. We prepared basic postures heuristically, and a final reaching posture is generated by modifying one of them. Heuristics, such as the fact that kneeling down is suitable for reaching near the ground, can be implemented easily by using this method. Methods that compose the reaching system, that is, basic posture selection, modification of postures for generating final reaching postures, balance compensation, footstep planning to realize desired feet position, and generation and execution of whole body motion to final reaching postures are described. Reaching to manually set positions and picking up a bat at various postures using visual information are shown as experiments to show the performance of the system.

One of the main challenges in the design of human-like walking robots (useful for service or entertainment applications as well as the study of human locomotion) is to obtain dynamic locomotion, as opposed to the static form of locomotion demonstrated by most of the current prototypes. A promising concept is the idea of passive dynamic walking; even completely unactuated and uncontrolled mechanisms can perform a stable gait when walking down a shallow slope. This concept enables the construction of dynamically walking prototypes that are simpler yet more natural in their motions than the static bipeds. This paper presents three additions to the concept of passive dynamic walking. First, hip actuation is added to increase the fore-aft stability and to provide power to the system, removing the need for a downhill floor. Second, a bisecting hip mechanism is introduced to allow the addition of a passive upper body without compromising the simplicity, efficiency and naturalness of the concept of passive dynamic walking. Third, skateboard-like ankle joints are implemented to provide 3D stability. These ankles couple the unstable sideways lean motion to yaw (steering), a kinematic coupling which provides sideways stability when walking with sufficient forward velocity. The three additions are investigated both with elementary simulation models and with prototype experiments. All three prototypes demonstrate an uncannily natural and stable gait while requiring only two foot switches and three on/off actuators.

In this paper, we address the problem of Navigation Among Movable Obstacles (NAMO): a practical extension to navigation for humanoids and other dexterous mobile robots. The robot is permitted to reconfigure the environment by moving obstacles and clearing free space for a path. This paper presents a resolution complete planner for a subclass of NAMO problems. Our planner takes advantage of the navigational structure through state-space decomposition and heuristic search. The planning complexity is reduced to the difficulty of the specific navigation task, rather than the dimensionality of the multi-object domain. We demonstrate real-time results for spaces that contain large numbers of movable obstacles. We also present a practical framework for single-agent search that can be used in algorithmic reasoning about this domain.

To synthesize whole-body behaviors interactively, multiple behavioral primitives need to be simultaneously controlled, including those that guarantee that the constraints imposed by the robot's structure and the external environment are satisfied. Behavioral primitives are entities for the control of various movement criteria, e.g. primitives describing the behavior of the center of gravity, the behaviors of the hands, legs, and head, the body attitude and posture, the constrained body parts such as joint-limits and contacts, etc. By aggregating multiple primitives, we synthesize whole-body behaviors. For safety and for efficient control, we establish a control hierarchy among behavioral primitives, which is exploited to establish control priorities among the different control categories, i.e. constraints, operational tasks, and postures. Constraints should always be guaranteed, while operational tasks should be accomplished without violating the acting constraints, and the posture should control the residual movement redundancy. In this paper, we will present a multi-level hierarchical control structure that allows the establishment of general priorities among behavioral primitives, and we will describe compliant control strategies for efficient control under contact interactions.

The mechanical and electrical system designs and system integration including controllers and sensory devices of the humanoid KHR-2 are presented. The design concept and the objective are also discussed. Since 2003, we have been developing KHR-2, which has 41 DOF (degrees of freedom). Each arm of KHR-2 has 11 DOF in total: 5 DOF/hand (i.e. fingers), 2 DOF/wrist, and 4 DOF/arm. Each leg constitutes 6 DOF. The head constitutes 6 DOF (2 DOF for eyes and 2 DOF at the neck), and the trunk has 1 DOF. KHR-2 has been mechanically designed to have a human friendly appearance and also wide ranges of angular motion. Its joint actuators have been designed in order to reduce motion uncertainties such as backlash. All axes of KHR-2 are under distributed control, which reduces the computational burden on the main controller and also to facilitate device expansions. We have developed a microprocessor-based sub-controller for servo motor operations, onto which sensory feedback is interfaced. The main controller, which is mounted on the back of the robot communicates with sub-controllers in real-time through CAN (Controller Area Network). Windows XP is used as the OS (Operating System), which enables rapid program development. RTX (Real Time eXtension) HAL extension software is used to realize the real-time control in the Windows XP environment. KHR-2 has several sensor types such as 3-axis F/T (Force/Torque) sensors at the foot and wrist, an inertia sensor system (accelerometer and rate gyro), and a CCD camera. The F/T sensor at the foot is crucially important for stable walking. The inertia sensor system is essential for determining the inclination of the robot with respect to the ground.

As humanoid robots become more commonplace in our society, it is important to understand the relation between humans and humanoid robots. In human face-to-face interaction, the observation of another individual performing an action facilitates the execution of a similar action, and interferes with the execution of a different action. This phenomenon has been explained by the existence of shared neural mechanisms for the execution and perception of actions, which would be automatically engaged by the perception of another individual's action. In one interference experiment, null interference was reported when subjects observed a robotic arm perform the incongruent task, suggesting that this effect may be specific to interacting with other humans. This experimental paradigm, designed to investigate motor interference in human interactions, was adapted to investigate how similar the implicit perception of a humanoid robot is to a human agent. Subjects performed rhythmic arm movements while observing either a human agent or humanoid robot performing either congruent or incongruent movements. The variance of the executed movements was used as a measure of the amount of interference in the movements. Both the human and humanoid agents produced a significant interference effect. These results suggest that observing the action of humanoid robots and human agents may rely on similar perceptual processes. Our findings suggest that experimental paradigms adopted from cognitive psychology can be used to derive measures for quantifying the degree of the implicit perception of a robot as human.