We provide a general approach for learning robotic motor skills from human demonstration. To represent an observed movement, non-linear differential equation is learned such that it reproduces this movement. Based on representation, we build library of movements by labeling each recorded movement according to task and context (e.g., grasping, placing, releasing). Our formulated generalization can be achieved simply adapting start goal parameter in the desired position values For object...

We review the work on data-driven grasp synthesis and methodologies for sampling ranking candidate grasps. divide approaches into three groups based whether they synthesize grasps known, familiar or unknown objects. This structure allows us to identify common object representations perceptual processes that facilitate employed technique. In case of known objects, we concentrate are recognition pose estimation. techniques use some form a similarity matching set previously encountered Finally...

In this paper, we present a new humanoid robot currently being developed for applications in human-centered environments. order robots to enter environments, it is indispensable equip them with manipulative, perceptive and communicative skills necessary real-time interaction the environment humans. The goal of our work provide reliable highly integrated platforms which on one hand allow implementation tests various research activities other realization service tasks household scenario. We...

Deep reinforcement learning (RL) methods generally engage in exploratory behavior through noise injection the action space. An alternative is to add directly agent's parameters, which can lead more consistent exploration and a richer set of behaviors. Methods such as evolutionary strategies use parameter perturbations, but discard all temporal structure process require significantly samples. Combining with traditional RL allows combine best both worlds. We demonstrate that off- on-policy...

Acquisition of new sensorimotor knowledge by imitation is a promising paradigm for robot learning. To be effective, action learning should not limited to direct replication movements obtained during training but must also enable the generation actions in situations has never encountered before. This paper describes methodology that enables generalization available knowledge. New are synthesized application statistical methods, where goal and other characteristics an utilized as queries...

We present a large-scale whole-body human motion database consisting of captured raw data as well the corresponding post-processed motions. This serves key element for wide variety research questions related e.g. to analysis, imitation learning, action recognition and generation in robotics. In contrast previous approaches, our considers motions observed subject objects with which is interacting. The information about human-object relations crucial proper understanding actions their...

This paper presents the resolve spatial constraints (RSC) algorithm for manipulation planning in a domain with movable obstacles. Empirically we show that our quickly generates plans simulated articulated robots highly nonlinear search space of exponential dimension. RSC is reverse-time samples future robot actions and constrains prior object displacements. To optimize efficiency RSC, identify methods sampling surfaces generating connecting paths between grasps placements. In addition to...

The increasing demand for robotic applications in dynamic unstructured environments is motivating the need dextrous end-effectors which can cope with wide variety of tasks and objects encountered these environments. human hand a very complex grasping tool that handle different sizes shapes. Many research activities have been carried out to develop artificial robot hands capabilities similar hand. In this paper mechanism design new humanoid-type (called TUAT/Karlsruhe Humanoid Hand)...

We propose a novel solution to the problem of inverse kinematics for redundant robotic manipulators purposes goal selection path planning. unify calculation configuration with searching in order avoid uncertainties inherent selecting configurations which may be unreachable because they currently lie components free space disconnected from initial configuration. adopt workspace heuristic functions that implicitly define regions and guide extension rapidly-exploring random trees (RRTs), are...

In this paper, we present efficient solutions for planning motions of dual-arm manipulation and re-grasping tasks. Motion such tasks on humanoid robots with a high number degrees freedom (DoF) requires computationally approaches to determine the robot's full joint configuration at given grasping position, i.e. solving Inverse Kinematics (IK) problem one or both hands robot. context, investigate inverse kinematics motion by combining gradient-descent approach in pre-computed reachability...

Linking human motion and natural language is of great interest for the generation semantic representations activities as well robot based on input. However, although there have been years research in this area, no standardized openly available data set exists to support development evaluation such systems. We, therefore, propose Karlsruhe Institute Technology (KIT) Motion-Language Dataset, which large, open, extensible. We aggregate from multiple capture databases include them our using a...

Having a representation of the capabilities robot is helpful when online queries, such as solving inverse kinematics (IK) problem for grasping tasks, must be processed efficiently in real world. When workspace representations, e.g. reachability an arm, are considered, additional quality information manipulability or self-distance can employed to enrich spatial data. In this work we present approach inverting precomputed representations order generate suitable base positions grasping....

In the recent past, recognition and localization of objects based on local point features has become a widely accepted utilized method. Among most popular are currently SIFT features, more SURF region-based such as MSER. For time-critical application object systems operating too slow (500-600 ms for images size 640×480 3 GHz CPU). The faster achieve computation time 150-240 ms, which is still active tracking or visual servoing applications. this paper, we present combination Harris corner...

The ability of computing a quality index for given configurations can be useful several applications in the context robotic manipulation. E.g. it used monitoring current state system or support decision processes, such as grasp selection humanoid robotics. Here, large set precomputed grasps object have to quickly filtered order select reachable sub set, which inverse kinematics (IK) problem has solved. In this work, we present an approach analyzing workspace capabilities manipulator store...

Large-scale human motion databases are key for research questions ranging from analysis and synthesis, biomechanics of motion, data-driven learning primitives, rehabilitation robotics to the design humanoid robots wearable such as exoskeletons. In this paper we present a large-scale database whole-body with methods tools, which allows unifying representation captured efficient search in database, well transfer subject-specific motions different embodiments. To end, is normalized regarding...

We present a data-driven, bottom-up, deep learning approach to robotic grasping of unknown objects using Deep Convolutional Neural Networks (DCNNs). The uses depth images the scene as its sole input for synthesis single-grasp solution during execution, adequately portraying robot's visual perception exploration scene. training consists precomputed high-quality grasps, generated by analytical grasp planners, accompanied with rendered objects. In contrast previous work on applying techniques...

A major goal of humanoid robotics is to enable safe and reliable human-robot collaboration in realworld scenarios. In this article, we present ARMAR-6, a new high-performance robot for various tasks, including but not limited grasping, mobile manipulation, integrated perception, bimanual collaboration, compliant-motion execution, natural language understanding. We describe how the requirements arising from these tasks influenced our design decisions, resulting vertical integration during...

Folding garments reliably and efficiently is a long standing challenge in robotic manipulation due to the complex dynamics high dimensional configuration space of garments. An intuitive approach initially manipulate garment canonical smooth before folding. In this work, we develop SpeedFolding, reliable efficient bimanual system, which given user-defined instructions as folding lines, manipulates an crumpled (1) smoothed (2) folded configuration. Our primary contribution novel neural network...

In this paper, we deal with imitation learning of arm movements in humanoid robots. Hidden Markov models (HMM) are used to generalize demonstrated a robot multiple times. They trained the characteristic features (key points) each demonstration. Using same HMM, key points that common all demonstrations identified; only those considered when reproducing movement. We also show how HMM can be detect temporal dependencies between both arms dual-arm tasks. created model human upper body simulate...