Keywords: Locomotion planning, Trajectory planning, Task planning
Abstract: In this paper, we propose a new approach to humanoid navigation through cluttered environments that exploits knowledge about different obstacle classes and selects appropriate robot actions. To classify objects from RGB images and decide whether an obstacle can be overcome by the robot with a corresponding action, e.g., by pushing or carrying it aside or stepping over or onto it, we train a convolutional neural network (CNN). Based on the associated action costs, we compute a cost grid of the environment on which a 2D path can be efficiently planned. This path encodes the necessary actions that need to be carried out to reach the goal. We implemented our framework in ROS and tested it in various scenarios with a Nao robot. As the experiments demonstrate, using the CNN the robot can robustly classify the observed obstacles into the different classes and exploit this information to efficiently compute solution paths. Our system finds paths also through regions where traditional planning methods are not able to calculate a solution or require substantially more time.

Keywords: Humanoid dynamics, Trajectory planning, Humanoids for human science
Abstract: Most of the existing stability control methods for humanoids aim at avoiding falling down, and it is difficult to cope with the sudden fall of robots. However, humanoids is an unstable system which can not avoid falling down. In this paper, we design a planning method of fall protection for humanoids according to the human fall motion. This method determines the contact position between the robot and the ground by adjusting the motion of the robot when it falls. In order to further reduce the damage to the robot, the appropriate cushioning material installed at the point of collision is selected to absorb the impact. The effectiveness of the proposed method is verified by BHR6P humanoid robot fall experiments.

Keywords: Multimodal perception, Deep Learning
Abstract: Robots play more and more important roles in our daily life. To better complete assigned tasks, it is necessary for the robots to have the ability to recognize their self-activities in real time. To perceive the environment, robots usually equipped with rich sensors, which can be used to recognize their self-activities. However, the intrinsics of the sensors such as accelerometer, servomotor and gyroscope may have significant differences, individual sensor usually exhibits weak performance in perceiving the environment. Therefore, multi-sensor fusion becomes a promising technique so that to achieve better performance. In this paper, facing the issue of robot self-activity recognition, we propose a framework to fuse information from multiple sensory streams. Our framework takes Recurrent Neural Network(RNN) that uses Long Short-Term Memory(LSTM) units to model temporal information conveyed in multiple sensory streams. In the architecture, a hierarchy structure is used to learn the sensor-specific features, a shared layer is used to fuse the features extracted from multiple sensory streams. We collect a dataset on PKU-HR6.0 robot to evaluate the proposed framework. The experiment results demonstrate the effectiveness of the proposed framework.

Attachments: Video Attachment
Keywords: System integration, Home, field, space, underwater, Visual perception
Abstract: Nowadays, studies have been conducted on humanoid robots with human mimetic structures.However, in the field of recognition, there is still much difference between humans and ordinary humanoid robots. For example, humans have movable eyes and use the degrees of freedom (DOF) of eyes in several ways: extension of the field of view, immediate changing of sight direction, focusing on objects to observe thoroughly. This DOF enables humans to adjust input images by aiming the direction of sight to desired objects, even when the body movement is limited.On the other hand, ordinary humanoid robots tend to have the cameras fixed to the head. Therefore, the humanoid has to move the head as a whole in order to change the sight direction of the camera.

In this study, we developed both a movable binocular high-resolution eye-camera unit, which is small enough to be installed in the humanoid head,and a system to change input images according to the environment.

The developed unit contains cameras, actuators, and motor control boards, being a stand-alone unit that enables the use of the movable eye. A small high-resolution auto-focus camera is used for this eye-camera unit. The developed system is used to adjust the images to the environment, controlling the recognition area by changing the direction of sight, size of input images and resolution.

Attachments: Video Attachment
Keywords: Novel sensing mechanisms, Tactile perception, Physical interaction
Abstract: Haptic sensation is an important modality for interacting with the real world. This paper proposes a general framework of inferring haptic forces on the surface of a 3D structure from internal deformations using a small number of physical sensors instead of employing dense sensor arrays. Using machine learning techniques, we optimize the sensor number and their placement and are able to obtain high- precision force inference for a robotic limb using as few as 9 sensors. For the optimal and sparse placement of the measurement units (strain gauges), we employ data-driven methods based on data obtained by finite element simulation. We compare data-driven approaches with model-based methods relying on geometric distance and information criteria such as Entropy and Mutual Information. We validate our approach on a modified limb of the “Poppy” robot and obtain 8 mm localization precision.

Attachments: Video Attachment
Keywords: Physical interaction, Novel sensing mechanisms, Tactile perception
Abstract: In this paper, we present developments of a real-time compliant motion control for a personal humanoid robot. Our approach allows to interpret and react to human guidance through touch using only joint encoders measurements to monitor contact direction and intensity on both static and moving links. This novel method is developed with consideration of minimal sensor requirement of the hardware platform to meet high affordability criteria. We demonstrate performances in experiments with a humanoid robot Pepper.

Keywords: Exoskeletons and assistive devices, Industrial
Abstract: In this paper, a soft upper-extremity passive robotic exosuit is presented. This exosuit intends to assist workers when they are carrying heavy loads. The feature of the proposed exosuit design is that it consists of a light plastic frame and a cable system mounted via joints on the frame. Polyethylene braid-style cables are used which have a high strength of extension up to 34kg and low deformability. Comparing to the rigid frame exoskeleton, the mass and the inertia of the cable are negligible. With the help of cable winding and locking mechanism, during the transporting phase of load carrying, the gravity of the load can be compensated by cables and redistributed on the shoulder and thigh. Hence the pressure of muscles on arm can be relieved. In order to have the optimal assistive effect, the positions of anchors and cable attachment points have been optimized. An experiment has been conducted with a preliminary prototype mounted on a mannequin test bench. Experiment result shows that with the help of this exosuit, the mannequin test bench can steadily hold its posture when a 10kg load is applied and meanwhile, its joint mobility is not affected when it is in the non-carrying-load mode.

Keywords: Locomotion, Body balancing
Abstract: Many methods exist for a bipedal robot to keep its balance while walking. In addition to step size and timing, other strategies are possible that influence the stability of the robot without interfering with the target direction and speed of locomotion. This paper introduces a multifaceted feedback controller that uses numerous different feedback mechanisms, collectively termed corrective actions, to stabilise a core keypoint-based gait. The feedback controller is experimentally effective, yet free of any physical model of the robot, very computationally inexpensive, and requires only a single 6-axis IMU sensor. Due to these low requirements, the approach is deemed to be highly portable between robots, and was specifically also designed to target lower cost robots that have suboptimal sensing, actuation and computational resources. The IMU data is used to estimate the yaw-independent tilt orientation of the robot, expressed in the so-called tilt phase space, and is the source of all feedback provided by the controller. Experimental validation is performed in simulation as well as on real robot hardware.

Keywords: Grasp and motion planning, Trajectory planning, Grasping and Manipulation
Abstract: Many methods have been developed for planning the motion of robotic arms for picking and placing, ranging from local optimization to global search techniques, which are effective for sparsely placed objects. Dense clutter, however, still adversely affects the success rate, computation times, and quality of solutions in many real-world setups. The current work integrates tools from existing methodologies and proposes a framework that achieves high success ratio in clutter with anytime performance. The idea is to first explore the lower dimensional end effector's task space efficiently by ignoring the arm, and build a discrete approximation of a navigation function, which guides the end effector towards the set of available grasps or object placements. This is performed online, without prior knowledge of the scene. Then, an informed sampling-based planner for the entire arm uses Jacobian-based steering to reach promising end effector poses given the task space guidance. While informed, the method is also comprehensive and allows the exploration of alternative paths over time if the task space guidance does not lead to a solution. This paper evaluates the proposed method against alternatives in picking or placing tasks among varying amounts of clutter for a variety of robotic manipulators with different end-effectors. The results suggest that the method reliably provides higher quality solution paths quicker, with a higher success rate relative to alternatives.

Attachments: Video Attachment
Keywords: Grasping and Manipulation, Tactile perception, Humanoid kinematics
Abstract: Grasping and manipulation are essential skills that humanoid robots need in order to operate in the human environment. Model-based methods require a precise calibration and suffer from high order non-linearity. While, neural-based representations does not require a dedicated calibration process to solve these tasks. However, some suffer from high generalization error that reduces the accuracy or require large-scale data collection. The role of sensory feedback is therefore important to adapt the action. We present a control framework to learn grasping with a soft gripper attached to a humanoid robot arm. The inverse kinematic model of the arm is acquired through motor babbling of a central pattern generator and encoded by a feed-forward neural network. To overcome the generalization error we provide the gripper with a tactile sensors array at each finger. The tactile servoing is used to correct the action before grasping. The proposed model has been tested in simulation, and on the real robot where a soft sensory gripper was used to interact with a human subject (Tactile Servoing). Successful grasping was achieved thanks to the integration of a learned inverse model with the sensory feedback.

Keywords: Sensorimotor learning, Skill modelling, Art, entertainment
Abstract: How do robots generalize the acquired motor representation in the workspace? In this paper, we present a framework that generates motor patterns for drawing arcs in the Cartesian workspace. The basic combinations of patterns resulting in a desired arc has been shown, where the patterns are generated by a Central Pattern Generator (CPG) model. The optimization of those combinations using Genetic Algorithm (GA) and then applying Inverse Distance Weighting (IDW) for generalization in the workspace are further discussed. However, due to the limitations of the aforementioned algorithms in the generalization of motor program, we proposed an approximation function using multilayer perceptron (MLP) to map the features of the trajectory of an arc into a corresponding motor parameters. After learning, we present scenarios, in which a humanoid robot, NAO, draws sketches in a 2D space. Unlike classical methods that use inverse kinematics to draw arcs through connecting several intermediate points in the Cartesian space, our proposed model generalizes the motor features of the pattern generator in the workspace.

Attachments: Video Attachment
Keywords: Grasp and motion planning, Grasping and Manipulation, Trajectory planning
Abstract: This paper presents a novel sampling-based motion planning method using bidirectional search with a time-configuration space representation that is able to efficiently generate collision-free trajectories in complex and non-static environments. Our approach exploits time indexing to separate a complex problem with mixed constraints into multiple sub-problems with simpler constraints that can be solved efficiently. We further introduce a planning framework by incorporating the proposed planning method enabling efficient pick-and-place of large objects in various scenarios. Simulation as well as hardware experiments show that the method also scales from redundant robot arms to mobile manipulators and humanoids. In particular, we have demonstrated that the proposed method is able to plan collision-free motion for a humanoid robot to pick up a large object placed inside a moving storage box while walking.

Keywords: Task planning
Abstract: Temporal stress is something that humans have to face nearly every day. Humans have to handle situations, where there is not much time left for a specific task. Most robotic systems, on the other hand, are not able to act in such temporally unstructured environments. For that reason, we present the novel Temporal Stressing Fast Downward (TSFD) planning system based on Temporal Fast Downward (TFD), which solves temporal problems using a modified heuristic forward search. With this planner, we introduce the novel concept of stressed actions for temporally bounded problems. Stressed actions enable a robot to accelerate or decelerate actions under consideration of an action-specific temporal cost function and the available time for plan execution. We further introduce an improved decision epoch search that allows complete planning with temporal gaps. Our evaluation in benchmark domains and on the real humanoid robot ARMAR-III shows that TSFD has the ability to produce plans of better makespan than TFD and is able to solve problems that could not be handled before. Furthermore, TSFD performs better in typical service robotics tasks than baseline approaches. Finally, we show that stressed actions greatly increase the possibility of finding feasible solutions in temporally bounded tasks.

Attachments: Video Attachment
Keywords: Deep Learning, Concept and strategy learning, Grasping and Manipulation
Abstract: We investigate the effectiveness of transfer learning for accelerating shallow and deep reinforcement learning of “mediated interaction tasks”. In these tasks, the desired effects cannot be created through direct interaction, but instead require the learner to discover how to exert suitable effects on the target object through involving a suitable “mediator object”. We focus on the case where transfer learning is applied to generalize experiences from source tasks that are solvable through direct, unmediated interaction, to target tasks that require mediated interaction for their solution. We find that transfer learning employed in this context leads to a significant acceleration of learning. A refinement of the basic transfer learning strategy that is motivated from the principle of scaffolding in psychology leads to further improvements, totalling to an overall speed-up factor of almost one order of magnitude for a reinforcement learner solving an “extension-of-reach” task in a 2D world with simulated physics.

Attachments: Video Attachment
Keywords: Humanoid kinematics, Body balancing
Abstract: Reducing the complexity of higher order problems can enable solving them in analytical ways. In this paper, we propose an analytic whole body motion generator for humanoid robots. Our approach targets inexpensive platforms that possess position controlled joints and have limited feedback capabilities. By analysing the mass distribution in a humanoid-like body, we find relations between limb movement and their respective CoM positions. A full pose of a humanoid robot is then described with five point-masses, with one attached to the trunk and the remaining four assigned to each limb. The weighted sum of these masses in combination with a contact point form an inverted pendulum. We then generate statically stable poses by specifying a desired upright pendulum orientation, and any desired trunk orientation. Limb and trunk placement strategies are utilised to meet the reference CoM position. A set of these poses is interpolated to achieve stable whole body motions. The approach is evaluated by performing several motions with an igus Humanoid Open Platform robot. We demonstrate the extendability of the approach by applying basic feedback mechanisms for disturbance rejection and tracking error minimisation.

Attachments: Video Attachment
Keywords: Skill modelling, Trajectory planning
Abstract: This paper introduces a novel online rolling over control technique based on energy concepts, to minimize the impact forces during the fall over of humanoids. To generate efficient rolling motion, critical parameters are defined by the insights drawn from a study on rolling, which are contact positions and attack angles. In addition, energy-injection velocity is proposed as an auxiliary rolling parameter to ensure sequential multiple contacts in rolling. An online rolling controller is synthesized to compute the optimal values of the rolling parameters. The first two parameters are to construct a polyhedron, by selection suitable contacts around the humanoid’s body. This polyhedron distributes the energy gradually across multiple contacts, thus called energy distribution polyhedron. The last parameter is to inject some additional energy into the system during the fall, to overcome energy drought and tip over successive contacts. The proposed controller, incorporated with energy minimization, distribution, and injection techniques, results in a rolling like motion and significantly reduces the impact forces, and it is verified in numerical experiments with a segmented planar model and a full humanoid model.

Attachments: Video Attachment
Keywords: Locomotion planning, Trajectory planning, Humanoid dynamics
Abstract: We present an optimization-based motion planning framework for producing dynamically rich and feasible motions for a 3D one-leg hopper in challenging terrains. We formulate dynamic motion planning as a nonlinear optimization problem that computes position and orientation of the centroidal model, position of the limb, contact forces, contact locations, and timings of the gait in one unified framework. The dynamics are represented as a single rigid body, while the equations of motion are derived using discrete mechanics with a variational quaternion-based integrator for the orientation. We validate the capabilities by planning complex motions in three challenging tasks: jumping over an obstacle, leaping over a gap, and performing a somersault. All contact forces generated by the proposed optimization are verified with accurate numerical simulation to prove the feasibility of the generated agile motions with respect to the kinematic, dynamic, and environmental constraints.

Keywords: Concept and strategy learning, Learning from demonstration, Grasping and Manipulation
Abstract: In recent optimization task-space controller, hierarchical task prioritization can be made strict or soft within a given level. Soft hierachization is made using task weighting. Yet the latter is not automated and weights are set ad-hoc. This empirical approach could be time-consuming and even leads to an infeasible result. During a specific episode in order to approximate the evolution of the weight of a task, we assign a Radial basis function network(RBFN) to each of the tasks. We use the Bayesian Optimization procedure to regulate the RBFNs corresponding to different tasks based on performances indexes that are extracted for a fixed episode. We benchmark the proposed solution with a dual-arm manipulation simulation where multiple potentially conflicting tasks are involved. First of all We can find that the proposed approach outperforms a hand-tuned controller in terms of tracking errors. In comparison with tuning the weights using another stochastic optimization technique, i.e. CMA-ES, we can find that the proposed approach requires much less samples to evaluate.

Attachments: Video Attachment
Keywords: Grasp and motion planning, Humanoid kinematics, Physical interaction
Abstract: Planning balanced whole-body reaching configurations is a fundamental problem in humanoid robotics on which manipulation and locomotion planners depend on. While finding valid whole-body configurations in free space and on flat terrains is relatively straightforward, the problem becomes extremely challenging when obstacle avoidance is taken into account, and when balancing on more complex terrains, such as inclined supports or steps. Previous work using Paired Forward-Inverse Dynamic Reachability Maps demonstrated fast end-pose planning on flat terrains at different heights by decomposing the kinematic structure and leveraging combinatorics. In this paper, we present an efficient whole-body end-pose planning framework capable of finding collision-free whole-body configurations in complex environments and on sloped support regions. The main contributions in this paper are twofold: (i) the integration of contact property information of support regions into both precomputation and online planning stages, including whole-body static equilibrium robustness, and (ii) the proposal of a more informed and meaningful sampling strategy for the lower-body. We focus on humanoid robots throughout the paper, but all the principles can be applied to legged platforms other than bipedal robots. We demonstrate our method on the NASA Valkyrie humanoid platform with 38 DoF over inclined supports. Analysis of the results indicate both higher success rates — greater than 95% and 80% on obstacle-free and highly cluttered environments, respectively — and shorter computation times compared to previous methods.

Keywords: Social interaction and acceptability, Modelling and simulating humans, Skill modelling
Abstract: The style of an action, i.e. the way it is performed, has a strong influence on interaction between humans. The very same gesture has very different consequences when it is performed aggressively or kindly, and humans are very sensitive to these subtle differences in others' behaviors. In this work we investigated how to endow a humanoid robot with behaviors expressing different vitality forms, by modulating robot action kinematics and voice intonation. Drawing inspiration from human voice and motion, we modified a passing action and a passing voice command performed by the robot to convey an aggressive or kind attitude. In a series of experiments we demonstrated that the humanoid was consistently perceived as aggressive or kind. Human behavior changed slightly in response to the different robot attitudes and was characterized by faster responses to robot utterances than to robot actions. The opportunity of humanoid behavior expressing vitality enriches the array of nonverbal communication that can be exploited by the robots to foster seamless interaction. Such behavior might be crucial in emergency and in authoritative situations in which the robot should instinctively be perceived as assertive and in charge, as in case of police robots or teachers.

Attachments: Video Attachment
Keywords: Task planning, Learning from demonstration, Grasping and Manipulation
Abstract: Solving high-level sequential decision tasks situated on physical robots is a challenging problem. Reinforcement learning, the standard paradigm for solving sequential decision problems, allows robots to learn directly from experience, but is ill-equipped to deal with issues of scalability and uncertainty introduced by real-world tasks. We reformulate the problem representation to better apply to robot manipulation using the relations of Object-Oriented MDPs (OO-MDPs) and the hierarchical structure provided by Abstract MDPs (AMDPs). We present a relation-based AMDP formulation for solving tabletop organizational packing tasks, as well as a demonstration-guided exploration algorithm for learning AMDP transition functions inspired by state- and action-centric learning from demonstration approaches. We evaluate our representation and learning methods in a simulated environment, showing that our hierarchical representation is suitable for solving complex tasks, and that our state- and action-centric exploration biasing methods are both effective and complementary for efficiently learning AMDP transition functions. We show that the learned policy can be transferred to different tabletop organizational packing tasks, and validate that the policy can be realized on a physical system.

Attachments: Video Attachment
Keywords: Locomotion planning, Body balancing, Locomotion
Abstract: This paper contributes towards the development and comparison of Divergent-Component-of-Motion (DCM) based control architectures for humanoid robot locomotion. More precisely, we present and compare several DCM based implementations of a three layer control architecture. From top to bottom, these three layers are here called: trajectory optimization, simplified model control, and whole-body QP control. All layers use the DCM concept to generate references for the layer below. For the simplified model control layer, we present and compare both instantaneous and Receding Horizon Control controllers. For the whole-body QP control layer, we present and compare controllers for position and velocity controlled robots. Experimental results are carried out on the one-meter-tall iCub humanoid robot. We show which implementation of the above control architecture allows the robot to achieve a walking velocity of 0.41 meters per second.

Keywords: Locomotion planning, Novel mechanism design
Abstract: In order to be applied in complex environment, humanoid robots are desired to have the ability of both biped walking and quadruped crawling. Crawling is a multi-contact motion. If the mechanism is completely rigid, there will be a closed kinematic chain of robot which is likely to cause damage to robot joints. Therefore, the robot needs to have some flexibility in mechanism. However, biped walking requires highly rigid mechanism to maintain walking stability. Consequently, it is a crucial issue to study crawling motion and biped walking motion in case of rigid-flexible mechanism. In this paper, firstly a 3-DoF rigid-flexible waist is proposed. The waist has rigidity when walking and flexibility when crawling. Then a crawling pattern generation algorithm based on CPG is presented, which solves the problem of difficult to plan crawling motion of robot with rigid-flexible mechanism. Finally, the validity of the proposed method is confirmed by experiments.

Attachments: Video Attachment
Keywords: Body balancing, Humanoid dynamics
Abstract: This paper presents a novel approach for friction compensation on humanoid robots originally designed for position control in order to enable torque-based control methods on such systems. Due to their design, this kind of robots lacks joint-torque sensors and is equipped with high-reduction gearboxes. Nevertheless, we can still apply torque commands using torque estimation from motor currents. Moreover, the high gear reduction ratio produces high dynamic friction which significantly affects the robot’s drives and must be taken into account. Considering the LuGre friction model, an adaptive friction compensator based on a second-order sliding mode is developed and illustrated on a 30-DoF humanoid robot with 86kg weight. The proposed method only relies on IMU data as well as joint position and velocity measurements from the joint encoders and is applied to the 12 DoFs of the robot's legs for CoM motion. The proposed control approach does not require the FT sensors mounted on the ankles.

Attachments: Video Attachment
Keywords: Humanoid dynamics, Locomotion
Abstract: This paper proposes a control strategy for the stabilization of a jet-powered flying humanoid robot. In particular, the contribution of the paper concerns the design of a control framework capable of tracking a desired robot position and orientation trajectory while flying. Asymptotic stability of the closed loop system is shown by means of a Lyapunov analysis. Simulations are carried out on a model of the humanoid robot iCub to verify the soundness of the proposed approach.

Attachments: Video Attachment
Keywords: Humanoid kinematics, Novel materials, Locomotion
Abstract: Bipedal robotic locomotion in granular media presents a unique set of challenges at the intersection of granular physics and robotic locomotion. In this paper, we perform a systematic experimental study in which biped robotic gaits for traversing a sandy slope are empirically designed using Zero Moment Point (ZMP) methods. We are able to implement gaits that allow our 7 degree-of-freedom planar walking robot to ascend slopes with inclines up to 10 degrees. Firstly, we identify a given set of kinematic parameters that meet the ZMP stability criterion for uphill walking at a given angle. We then find that further relating the step lengths and center of mass heights to specific slope angles through an interpolated fit allows for significantly improved success rates when ascending a sandy slope. Our results provide increased insight into the design, sensitivity and robustness of gaits on granular material, and the kinematic changes necessary for stable locomotion on complex media.

Attachments: Video Attachment
Keywords: System integration, Visual perception
Abstract: KinectFusion is an impressive algorithm that was introduced in 2011 to simultaneously track the movement of a depth camera in the 3D space and densely reconstruct the environment as a Truncated Signed Distance Formula (TSDF) volume, in real-time. In 2012, we introduced the Moving Volume KinectFusion method that allows the volume/camera move freely in the space. In this work, we further develop the Moving Volume KinectFusion method (as rxKinFu) to fit better to robotic and perception applications, especially for locomotion and manipulation tasks. We describe methods to raycast point clouds from the volume using virtual cameras, and use the point clouds for heightmaps generation (e.g., useful for locomotion) or object dense point cloud extraction (e.g., useful for manipulation). Moreover, we present different methods for keeping the camera fixed with respect to the moving volume, fusing also IMU data and the camera heading/velocity estimation. Last, we integrate and show some demonstrations of rxKinFu on the mini-bipedal robot RPBP, our wheeled quadrupedal robot CENTAURO, and the newly developed full-size humanoid robot COMAN+. We release the code as an open-source package, using the Robotic Operating System (ROS) and the Point Cloud Library (PCL).

Keywords: Locomotion planning, Humanoid dynamics
Abstract: Energy efficiency of biped walking is an crucial topic for humanoid robot’s research. Rapid computing is also important for online planning and model transplantation. Many dynamic models for characterizing humanoids’ walking have been developed, such as conventional 3 dimensional inverted pendulum (IPM), linear inverted pendulum (LIPM). This paper proposed an improved inverted pendulum model constrained on cylindrical surface (CIPM), combining the advantages of computing and energy efficiency for humanoids’ walking planning. Walking patterns with different speeds can be generated by CIPM. The constraint of cylindrical surface results in low coupling between displacement variables for tested robot and the energy consumption is less than that generated based on LIPM. The advantages of CIPM over IPM and LIPM were proved by mathematic analysis, simulations of bipedal walking with different speeds.

Attachments: Video Attachment
Keywords: Body balancing, Humanoid dynamics
Abstract: Any arbitrary motion generated by a humanoid robot produces a yaw moment which may exceed the one created by the friction between its feet and the ground, inducing a yaw rotation that deviates the robot from its desired path. This paper describes an on-line compensation scheme for the yaw moment of a humanoid robot about the Zero Moment Point (ZMP), formulated as a task of a Quadratic Program (QP) solving for multiple weighted objectives and constraints. This allows to use the motion of every single link of the robot to contribute to the compensation, according to the relative weight of other primary tasks that it should perform. Within the proposed approach the yaw moment is partially compensated; that is, mostly when exceeding a predefined threshold, allowing to slow down the residual motion of the links triggered by the compensation.

Attachments: Video Attachment
Keywords: Locomotion, Locomotion planning, System integration
Abstract: We developed a robot which can do both of active walking (all joints are actively controlled by actuators) and semi-passive walking (hip joints are passive and spring attached). In this paper, we summarize three technologies to achieve the development. The first one is small and high-strength clutch mechanism to sustain massive weight of life-size robot. The second one is semi-passive walk controller to consider passive joint dynamics. The last one is model parameter identification considering not only body parameters but also environment ones such as ground slope to achieve unstable motion similar to simulated result in real world.

Keywords: Learning from demonstration, Trajectory planning
Abstract: In the context of humanoid skill learning, movement primitives have gained much attention because of their compact representation and convenient combination with a myriad of optimization approaches. Among them, a well- known scheme is to use Dynamic Movement Primitives (DMPs) with reinforcement learning (RL) algorithms. While various remarkable results have been reported, skill learning with the physical constraints has not been sufficiently investigated. For example, when RL is employed to optimize the robot joint trajectories, the exploration noises could make the resulting trajectory out of the joint limits. In this paper, we focus on robot skill learning featured by joint limit avoidance, where the novel Constrained Dynamic Movement Primitives (CDMPs) are developed. By using the exogenous states instead of the original DMPs states, CDMPs are capable of maintaining the DMPs within the joint limits. We validate CDMPs on the humanoid robot iCub, showing the applicability of our approach.

Attachments: Video Attachment
Keywords: Locomotion, Humanoid dynamics, Humanoid kinematics
Abstract: Most humanoid robots still walk with bent knees and flat feet which is considered highly unnatural, i.e. not biologically inspired, and also energy inefficient. The paradigm and benefits of walking with non-bent knees and with an active toe joint are explored in this study. Non-bent knee walking trajectories are created using an instantaneous capture point (ICP) planner within a momentum based quadratic program (QP) whole body control framework. The toe joint trajectories are obtained as an emergent behavior of the QP determined by under-constraining the objective function and modeling movement of the toe joint as a torsional spring. A comparison between similar systems with and without toe joints reveal a stronger thrust vector during toe-off, reduced knee joint angles and a more human like gait. Experiments in simulation are conducted on the Atlas humanoid robot. Keywords: Humanoid Robots, Toe Joint, Non-bent knees.

Attachments: Video Attachment
Keywords: Humanoid ethics and philosophy, Social interaction and acceptability, Multimodal interaction
Abstract: Lie detection is a necessary skill for a variety of social professions, including teachers, reporters, therapists, and law enforcement officers. Autonomous system and robots should acquire such skill to support professionals in numerous working contexts. Inspired by literature on human-human interaction, this work investigates whether the behavioral cues associated to lying – including eye movements and response temporal features – are apparent also during human-humanoid interaction and can be leveraged by the robot to detect deception. The results highlight strong similarities in the lying behavior toward humans and the robot. Further, the study proposes an implementation of a machine learning algorithm that can detect lies with an accuracy of 75%, when trained with a dataset collected during human-human and human robot interaction. Consequently, this work proposes a technological solution for humanoid interviewers that can be trained with knowledge about lie detection and reuse it to counteract deception.

Attachments: Video Attachment
Keywords: Grasping and Manipulation, Grasp and motion planning
Abstract: When a humanoid robot carries an object, it should recognize surroundings to avoid obstacles while holding the object stably. Previous methods usually hold an object in front of its body. However, it causes visual occlusions and instability of holding when the object is supported by only its hands. To solve these problems, we propose methods for making a humanoid robot lift and carry an object on the head. By holding an object on the head, a robot can recognize its surroundings easily as the object goes out of its sight. In addition, as the object is supported by both hands and the head, three points in total, carriage becomes more stabilized. We also implement our methods on the humanoid robot JAXON and show the methods can be applied to the real robot.

Attachments: Video Attachment
Keywords: Energy system for humanoids, System integration, Novel actuation mechanisms
Abstract: In this paper, we present the design of a tiny high-power motor driver without a liquid cooling system. A high-power humanoid robot JAXON developed in our laboratory has a liquid cooling system for cooling motors and motor drivers. Thanks to the liquid cooling system, JAXON realizes high-power motion. However, the liquid cooling system makes the robot heavy. Hence, we designed a motor driver which is air cooling and smaller than the existing one while keeping its performance. There are two key points to realize tiny high-power motor driver. One is absorbing temperature rise due to instantaneous high current by using a heavy copper circuit board as a thermal buffer. The other is keeping steady board temperature low by reducing power consumption of motor driver. By applying developed motor driver to JAXON1, JAXON1 was able to reduce the weight of 7.1 [kg].

Attachments: Video Attachment
Keywords: Locomotion, Locomotion planning, Humanoid kinematics
Abstract: This paper explores two new configurations for humanoid robot balancing and locomotion. Centroidal momentum manipulability analysis has been performed to study the features of the newly proposed configurations. Numerical simulations show that they outperform the regular ones in terms of angular momentum manipulability. More than that, the new configurations allow the humanoid robot to perform cross-step motions which is usually risky or mechanically impossible for most existing robots. However, cross-step introduces non-convex feasible region which makes it difficult to be incorporated into our existing step planner. Therefore, a simple heuristic has been proposed to help choosing a sub-convex region for the step planner. To validate the cross-step movement, walking simulations have been performed.

Keywords: Humanoid dynamics, Locomotion planning
Abstract: It is very easy for biped robots to fall down. Some previous studies have been carried out to detect the fall state and protect the robot from damage. But it is not enough to detect a fall. It is very important for the biped robot to predict whether it will fall in the future based on the current state. In this paper, we consider a fall state predicted problem for bipedal robots. Based on the D 'Alembert principle, this method can predict the fall state at the moment the biped robot deviates from the normal state in every conditions such as standing and walking. It can give the robot more time to recover from the unstable state or protect itself from damage. And its stable control strategy matching the proposed method is also proposed to protect the robot from falling. The result is verified via simulations.

Keywords: Novel sensing mechanisms, Software and hardware architecture
Abstract: Robots that perform useful heavy-duty tasks are gaining attention in the field of construction, mining, and disaster recovery. For robust accomplishment of such tasks, control of interaction force is important fundamental functionality. Use of joint torque sensors is the most common method for robots that realize physical interaction. However, torque sensors add weight and reduce joint stiffness which result in loss of mobility performance. In this paper, joint torque sensing using link structure strain measurement is presented. Redundant strain gauges, placed in unstructured manner, are used to measure link deformation, which are then used to estimate all 6 components of the wrench acting on a link structure. Joint torque is then extracted from this wrench, which minimizes the cross-talks of the force measurement. Redundancy enhances the measurement accuracy and realizes fault tolerant force measurement. Simulation and experimental results of the measurement concept together with the fault recovery method are presented.

Attachments: Video Attachment
Keywords: Sensorimotor learning
Abstract: The prohibitively amount of data required when learning complex nonlinear policies, such as deep neural networks, has been significantly reduced with guided policy search (GPS) algorithms. However, while learning the control policy, the robot might fail and therefore generate unacceptable guiding samples. Failures may arise, for example, as a consequence of modeling or environmental uncertainties, and thus unsuccessful interactions should be explicitly considered while learning a complex policy. Currently, GPS methods updates the robot policy discarding or giving low probability to unsuccessful trials. In other words, they overlook the existence of poorly performing executions, and therefore tend to underestimate the information of these interactions in next iterations. In this paper we propose to learn deep neural network controllers with an extension of GPS that considers trajectories optimized with dualist constraints. These constraints are aimed at assisting the policy learning so that the trajectory distributions updated at each iteration are similar to good trajectory distributions while differing from bad trajectory distributions. We show that neural network policies guided by trajectories optimized with our method reduce the failures during the policy exploration phase, and therefore encourage safer interactions. This may have a relevant impact in tasks that involve physical contact with the environment or human partners in collaborative scenarios.

Keywords: Exoskeletons and assistive devices, Prostheses & Ortheses, Medical, health and mental care
Abstract: Cable-driven rehabilitation devices (CDRDs) have been studied by many researchers in the past decade. While a CDRD rotates a human joint by generating an assistant moment about the axis of the joint, it also creates a resultant force acting on the joint as long as the assistant moment is nonzero. Such a joint force may cause excessive joint wear or even break the joint if it exceeds a threshold. Thus, it is critical to analyze not only the assistant moment generated by a CDRD to rotate the joint, but also the joint force to have a safe and effective rehabilitation training. This paper studies how a CDRD with three degrees of freedom (DOFs) and four cables exerts a joint force on a general three-DOF human joint. The kinematics and dynamics models of the CDRD are established and the joint force needed to provide the assistant moment is derived mathematically at first. Then, an index to evaluate the efficiency of a CDRD in providing assistant moment (i.e., moment efficiency) is proposed. Lastly, a case study of the flexion and extension of the knee assisted by a CDRD is presented to demonstrate the derivation of the joint force and the usage of the moment efficiency index. The moment efficiency index not only promotes the safety of rehabilitation training, but also provides a guideline for the design of CDRDs

Attachments: Video Attachment
Keywords: Grasping and Manipulation, Teleoperation, Novel mechanism design
Abstract: This paper presents a series of control strategies for soft compliant manipulators. We provide a novel approach to control multi-fingered tendon-driven foam hands using a CyberGlove and a simple ridge regression model. The results achieved include complex posing, dexterous grasping and in-hand manipulations. To enable efficient data sampling and a more intuitive design process of foam robots, we implement and evaluate a finite element based simulation. The accuracy of this model is evaluated using a Vicon motion capture system. We then use this simulation to solve inverse kinematics and compare the performance of supervised learning, reinforcement learning, nearest neighbor and linear ridge regression methods in terms of their accuracy and sample efficiency.

Attachments: Video Attachment
Keywords: Locomotion, Body balancing, Humanoid dynamics
Abstract: In this paper, we propose the concept of enhanced capture point. Then we develop the enhanced CP controller with two enhanced CP control methods eCPS, eCPT. Their stability and disturbance rejection ability are analyzed. In addition, we introduce the concept of controllable region of CP and compare the enhanced CP control with CP control on the disturbance rejection performance. The simulation results of the linear inverted pendulum model verify the effectiveness of the proposed method, and the use of eCPT to achieve a more stable gait of the robot.

Keywords: Cognitive development, Concept and strategy learning, Multimodal perception
Abstract: Robots are progressively moving into spaces that have been primarily shaped by human agency; they collaborate with human users in different tasks that require them to understand human language so as to behave appropriately in space. To this end, a stubborn challenge that we address in this paper is inferring the syntactic structure of language, which embraces grounding parts of speech (e.g., nouns, verbs, and prepositions) through visual perception, and induction of Combinatory Categorial Grammar (CCG) in situated human-robot interaction. This could pave the way towards making a robot able to understand the syntactic relationships between words (i.e., understand phrases), and consequently the meaning of human instructions during interaction, which is a future scope of this current study.

Keywords: Humanoid dynamics, Humanoid kinematics, Sensorimotor learning
Abstract: Robot inverse dynamics modeling is performed mainly via standard system identification and/or machine learning techniques. In this paper we part from the theoretical framework of First-Order Principles Networks (FOPnet), combining data-aided learning with basic knowledge to learn the model of a targeted robot. The framework, previously used for learning the dynamics of a fixed-base serial manipulator, is now extended to the learning of the kinematics and dynamics of tree-structured robots with floating base. Our approach leverages the principle of compositionality to separate the main problem into two partially independent modules. The first defines the robot's body schema by characterizing its morphology and topology. The second is dependent upon the latter and defines the inertial properties of the multi-body system. To demonstrate the capabilities of the approach, a simulated humanoid robot with 30 degrees of freedom is used. We discuss the implementation of our method and evaluate its estimation and generalization capabilities in comparison with other common machine learning approaches. Finally, we present experimental results on a 7-DoF manipulator.

Attachments: Video Attachment
Keywords: Humanoids for human science, Modelling and simulating humans
Abstract: Tendon-driven musculoskeletal humanoids typically have complex structures similar to those of human beings, such as ball joints and the scapula, in which encoders cannot be installed. Therefore, joint angles cannot be directly obtained and need to be estimated using the changes in muscle lengths. In previous studies, methods using table-search and extended kalman filter have been developed. These methods express the joint-muscle mapping, which is the nonlinear relationship between joint angles and muscle lengths, by using a data table, polynomials, or a neural network. However, due to computational complexity, these methods cannot consider the effects of polyarticular muscles. In this study, considering the limitation of the computational cost, we reduce unnecessary degrees of freedom, divide joints and muscles into several groups, and formulate a joint angle estimation method that takes into account polyarticular muscles. Also, we extend the estimation method to propose a joint angle estimation method using only the relative changes in muscle lengths. By this extension, which does not use absolute muscle lengths, we do not need to execute a difficult calibration of muscle lengths for tendon-driven musculoskeletal humanoids. Finally, we conduct experiments in simulation and actual environments, and verify the effectiveness of this study.

Attachments: Video Attachment
Keywords: System integration, Grasp and motion planning, Visual perception
Abstract: Autonomous dual-arm manipulation is an essential skill to deploy robots in unstructured scenarios. However, this is a challenging undertaking, particularly in terms of perception and planning. Unstructured scenarios are full of objects with different shapes and appearances that have to be grasped in a very specific manner so they can be functionally used. In this paper we present an integrated approach to perform dual-arm pick tasks autonomously. Our method consists of semantic segmentation, object pose estimation, deformable model registration, grasp planning and arm trajectory optimization. The entire pipeline can be executed on-board and is suitable for on-line grasping scenarios. For this, our approach makes use of accumulated knowledge expressed as convolutional neural network models and low-dimensional latent shape spaces. For manipulating objects, we propose a stochastic trajectory optimization that includes a kinematic chain closure constraint. Evaluation in simulation and on the real robot corroborates the feasibility and applicability of the proposed methods on a task of picking up unknown watering cans and drills using both arms.

Attachments: Video Attachment
Keywords: Novel mechanism design, Novel actuation mechanisms, Humanoid kinematics
Abstract: This paper reports design and implementation of a multiloop robotic eyeball mechanism that enables synchronously rotating two eyeballs using a single actuator. Our optimally designed eyeball mechanism can help in implementing light weight, agile and energy efficient robotic heads. To the best of our knowledge, no existing eyeball mechanism is able to synchronously rotate two eyeballs using a single actuator. This work demonstrates use of a multiloop mechanism for reducing the number of required actuators and hence reducing the overall power consumption. Our eyeball mechanism incorporates an optimally designed 4-PS (prismatic-spherical) plus 1-P (passive support) construct. This partially passive construct comprises of a double-dwell end-cam plus a 4-follower arrangement. The cam-follower arrangement also augments a control strategy for synchronously rotating eyeballs and irides. We also present a methodology for determining the position kinematics of this 4-degree of freedom robotic eyeball mechanism.