Reactive Reaching and Grasping on a Humanoid

Towards Closing the Action-Perception Loop on the iCub

urgen Leitner, Mikhail Frank, Alexander F

orster and J

urgen Schmidhuber

Dalle Molle Institute for Artiﬁcial Intelligence (IDSIA) and Scuola universitaria professionale della Svizzera italiana

(SUPSI) and Universit

a della Svizzera italiana (USI), Lugano, Switzerland

Keywords:

Humanoid Robot, Robot Control, Object Manipulation, Reactive Control, Collision Avoidance, Robot

perception, Eye-hand Coordination, Computer Vision.

Abstract:

We propose a system incorporating a tight integration between computer vision and robot control modules on

a complex, high-DOF humanoid robot. Its functionality is showcased by having our iCub humanoid robot

pick-up objects from a table in front of it. An important feature is that the system can avoid obstacles – other

objects detected in the visual stream – while reaching for the intended target object. Our integration also

allows for non-static environments, i.e. the reaching is adapted on-the-ﬂy from the visual feedback received,

e.g. when an obstacle is moved into the trajectory. Furthermore we show that this system can be used both in

autonomous and tele-operation scenarios.

1 INTRODUCTION

After centuries of imagining automated machines that

would help humans, the last few decades brought an

ever growing number of these programmable appli-

ances into existence. While there are nowadays thou-

sands of robotic arms fulﬁlling automation tasks in

factories, robots are not yet in wide-spread use outside

industrial applications. More and more focus is put

on enabling other ares of utilisation, e.g. the use of

robots in household or healthcare settings. For robotic

systems to perform tasks in such environments, which

are not speciﬁcally designed for robots, a better un-

derstanding – i.e. perception – of the situation and a

closer coordination between action and perception –

i.e. adapting to changes occurring in the scene – are

required (Ambrose et al., 2012; Kragic and Vincze,

2009; Kemp et al., 2007).

The goal of this work is to extend the capabili-

ties of our robot to allow for more autonomous and

more adaptive – some would say, more ‘intelligent’ –

behaviours. Our platform is a state-of-the-art, high

degree-of-freedom (DOF) humanoid (see Figure 1)

and we focus on object manipulation in non-static en-

vironments. To facilitate this a close integration of

computer vision and control modules was developed

and implemented. Drawing inspiration from a wide

variety of techniques, our system incorporates, e.g.,

machine learning approaches to implement vision.

The paper revisits some of the previous ap-

proaches in the ﬁelds of robot vision and control, as

well as previous work on building integrated robotic

systems using perception and action subsystems in

Section 2. We then describe our approach to build-

ing a tightly coupled framework for visually-guided

reaching and grasping in detail (Section 3), before we

present our high DOF, state-of-the-art robot platform

(Section 4). Section 5 explains the experiments and

our ﬁndings, while conclusions (Section 6) and plans

for future research are found at the end.

2 BACKGROUND

Although there exists a rich body of literature in

computer vision, path planning, and feedback con-

trol, wherein many critical subproblems are addressed

Figure 1: Our research platform, the iCub humanoid robot.

102

Leitner J., Frank M., Förster A. and Schmidhuber J..

Reactive Reaching and Grasping on a Humanoid - Towards Closing the Action-Perception Loop on the iCub.

DOI: 10.5220/0005113401020109

In Proceedings of the 11th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2014), pages 102-109

ISBN: 978-989-758-039-0

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

individually, most demonstrable behaviours for hu-

manoid robots do not effectively integrate elements

from all three disciplines. Consequently, tasks that

seem trivial to us humans, such as picking up a spe-

ciﬁc object in a cluttered environment, remain beyond

the state-of-the-art in experimental robotics.

Sensory feedback is of critical importance to de-

cide and act in unstructured environments. Robots re-

quire the ability to perceive their surroundings, as it

is generally not feasible or possible to provide all the

information needed a priori (the most striking exam-

ple being robotic space exploration, where only little

about the environment the robot is sent to is known

before each mission).

2.1 Perception and Robot Vision

Creating a functional perception system is a hard, but

important, problem. It is a prerequisite for robots to

act in a purposeful, ‘intelligent’ way.

A wide range of sensors have been used to build

models of the environment. Visual feedback tends

to be the most interesting and active research area,

as cameras tend to be cheap and in widespread use

nowadays. It has extensively been used in mobile

robot applications, for obstacle avoidance, mapping

and localisation (Davison and Murray, 2002; Karls-

son et al., 2005). Furthermore there exists a natu-

ral tendency to be inspired by human perception, in

which vision plays a primary role, and behaviour, es-

pecially when working with humanoid robots. The

ﬁeld of Computer Vision spawned a multitude of re-

search sub-domains, such as, scene reconstruction,

event detection, object recognition, motion estima-

tion, etc. An important problem in robot vision is

that of determining whether the image data contains

some speciﬁc object, feature, or activity. The inter-

ested reader is referred to a nice survey of the vari-

ous approaches (Cipolla et al., 2010). This has been

researched for quite some time but the task seems

harder than expected. No solution for the general

case, i.e. detecting arbitrary objects in arbitrary sit-

uations, exists (Kemp et al., 2007). Problems arise

because of real-life environmental impacts, such as,

changing light conditions, incomplete or inconsistent

data acquisition (Kragic and Vincze, 2009) (see also

Figure 2). Most object detection applications are us-

ing SIFT (Lowe, 1999) or SURF (Bay et al., 2006)

with extensions for higher robustness (St

uckler et al.,

2013). For real time object tracking in videos frames

contour trackers are often used (Panin et al., 2006).

Experimental robotics though still relies heavily on

artiﬁcial landmarks and ﬁducial markers to simplify

(and speed-up) the detection problem.

Once an object is successfully detected in the cam-

era images, it is important to localise it, so it can

be placed into a model of the environment. ‘Spa-

tial Perception’, as this is known, develops in humans

over time from observation and interaction with the

world. Research in brain- and neuro-science show

clear trends on what changes during this development,

but how these changes happen is not clear (Plumert

and Spencer, 2007).

To obtain a distance measure stereo vision systems

use two images taken from different angles (Hartley

and Zisserman, 2000). In humanoid robot scenarios

a method for localisation is required, which is able to

cope with motion in the robot’s head, gaze and upper

body. While projective geometry approaches work

well under carefully controlled experimental circum-

stances, they are not easily transferred to robotics ap-

plications though (as seen by the out-of-the-box lo-

calisation capabilities of the iCub (Pattacini, 2011)).

It was previously shown that, even on complex hu-

manoid robots, this stereo localisation can be learned

(Leitner et al., 2012c).

2.2 Motion and Actions

Only after the robot knows which objects are in the

environment and where they are located can it start to

interact with them. The Path Planning Problem is an

important problem in robot motion research. It de-

notes the issue to ﬁnd motions that pursue a set goal,

e.g. reaching, while deliberately avoiding obstacles

and other constraints. Solving this problem is critical

to deploying complex robots in unstructured environ-

ments. An overview of interesting approaches to mo-

tion planning can be found in the textbook “Planning

Algorithms” (LaValle, 2006). To solve the path plan-

ning problem is generally expensive, therefore robots

controlled this way are typically very deliberate and

Figure 2: Detection of complex objects, e.g. a tea box, in

changing poses, different light and when partially occluded

is a hard problem in robot vision.

ReactiveReachingandGraspingonaHumanoid-TowardsClosingtheAction-PerceptionLoopontheiCub

103

slow. This is often perceived during robotics demos

(e.g. at recent DARPA Robotics Challenge Trials),

where one can see them ‘think’ ﬁrst, often for quite

a bit, before a motion is executed.

A more reactive approach, instead of the ‘think

ﬁrst, act later’ paradigm, emerged in the 1980s. A

variety of approaches have been applied to quickly

generate control commands, without searching the

robot’s conﬁguration space (Khatib, 1986; Brooks,

1991; Schoner and Dose, 1992). All these use local

information from the workspace, and transform it into

motor commands according to some heuristics. It is

therefore not surprising that these approaches excel at

fast, reactive obstacle avoidance while they have trou-

ble with global planning tasks. Because of this such

approaches have become popular in the context of

safety and human-robot interaction (De Santis et al.,

2007; Dietrich et al., 2011).

Once a motion is found to reach for a certain ob-

ject the actual manipulation is not a trivial thing, even

for humans. For example, research shows that even a

prototypical precision grasp task is not matured until

the age of 8-10 years (Forssberg et al., 1991)). Recent

results though highlight the progress and advanced

state of research in robot grasping (Carbone, 2013).

2.3 Integration

To allow for a variety of objects to be picked up from

various positions the robot needs to see, act and re-

act within an integrated control system. For example,

methods enabling a 5 DOF robotic arm to pick up ob-

jects using a point-cloud generated model of the world

and objects are available to calculate reach and grasp

behaviours (Saxena et al., 2008). In 2010 a technique

for robots to pick up non-rigid objects, such as, tow-

els was presented (Maitin-Shepard et al., 2010). It

allows to reliably and robustly pick up a towel from a

table by going through a sequence of vision-based re-

grasps and manipulations-partially in the air, partially

on the table.

Even when sufﬁcient manipulation skills are avail-

able these need to be constantly adapted by an

perception-action loop to yield desired results. A

must read for roboticists is ‘Robotics, Vision and

Control’ (Corke, 2011)’. It puts the close integra-

tion of these three components into the spotlight and

describes common pitfalls and issues when trying to

build such systems with high levels of sensorimotor

integration. In the DARPA ARM project, which aims

to create highly autonomous manipulators capable of

serving multiple purposes across a wide variety of ap-

plications, the winning team showed an end-to-end

system that allows the robot to grasp and pick-up di-

verse objects (e.g. a power drill, keys, screwdrivers,

etc.) from a table by combining touch and LASER

sensing (Hudson et al., 2012).

Visual Servoing (Chaumette and Hutchinson,

2006) is a commonly used approach to closed-loop vi-

sion based control of robotic systems – i.e. some level

of hand-eye coordination. It has been shown to work

as a functional strategy to control robots without any

prior hand-eye calibration (Vahrenkamp et al., 2008).

3 CLOSING THE

ACTION-PERCEPTION LOOP

Our aim is to generate a pick-and-place operation

for the iCub. For this, functional motion and vision

subsystems are integrated to create a closed action-

perception loop. The vision-side detects and localises

the object continuously, while the motor-side tries to

reach for target objects avoiding obstacles at the same

time. A grasping of the object is triggered when the

hand is near the target.

3.1 Action Side: MoBeE

Modular Behavioral Environment (MoBeE) (Frank

et al., 2012) is a software infrastructure to realise

complex, autonomous, adaptive and foremost safe

robot behaviours. It acts as an intermediary between

three loosely coupled types of modules: the Sen-

sor, the Agent and the Controller. These correspond

to abstract solutions to problems in Computer Vi-

sion, Motion Planning, and Feedback Control, respec-

tively. An overview of the system is depicted in Fig-

ure 3. The framework is robot independent, and can

exploit any device that controlled via YARP (Metta

et al., 2006). It also supports multiple interacting

robots, and behavioural components are portable and

reusable thanks to their weak coupling. MoBeE con-

trols the robot constantly, according to the following

second order dynamical system:

M ¨q(t) +C ˙q(t) + K(q(t) − q

∗

) =

∑

(t) (1)

where q(t) ∈ R

is the vector function representing

the robot’s conﬁguration, M, C, K are matrices con-

taining mass, damping and spring constants respec-

tively. q

∗

denotes an attractor (resting pose) in con-

ﬁguration space. Constraints on the system are im-

plemented by forcing the system via f

(t), providing

automatic avoidance of kinematic infeasibilites aris-

ing from joint limits, cable lengths, and collisions.

An agent can interact with MoBeE, instead of di-

rectly with the robot, by sending arbitrary high-level

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

104

Figure 3: The Modular Behavioral Environment Architec-

ture: MoBeE implements low-level control and enforces

necessary constraints to keep the robot safe and operational

in real-time. Agents (left) are able to send high-level com-

mands, while a kinematic model (top) is driven by the

stream of encoder positions (right). The model computes

ﬁctitious constraint forces, which repel the robot from col-

lisions, joint limits, and other infeasibilities. These forces,

(t), are passed to the controller (middle), which computes

the attractor dynamics that governs the actual movement of

the robot.

control commands. For example, when a new at-

tractor q

∗

is set to a desired pose by an agent, e.g.

by calculating the inverse kinematics of an opera-

tional space point, q(t) begins to move toward q

∗

The action then terminates either when the dynamical

system settles or when a timeout occurs, depending

on the constraint forces f

(t) encountered during the

transient response.

3.2 Perception Side: icVision & CGP-IP

icVision (Leitner et al., 2012b) is an open-source

framework consisting of distributed YARP modules

performing computer vision related tasks in support

of cognitive robotics research (Figure 4). It includes

the modules for the detection and identiﬁcation of ob-

jects (in the camera images, referred to as Filters), as

well as the localisation of the objects in the robot’s

operational space (3D Cartesian space). At the centre

is the icVision core module, which handles the con-

nection with the hardware and provides housekeeping

functionality (e.g., extra information about the mod-

ules started and methods to stop them). Currently

available modules include object detection, 3D lo-

calisation, gazing control (attention mechanism) and

saliency maps. Standardised interfaces allow for easy

swapping and reuse of modules.

The main part in object detection, the binary seg-

mentation of the object from the background (see Fig-

ure 4 on the right), in the visual space, is performed in

separate icVision ﬁlter modules. Each one is trained

Code available at: https://github.com/Juxi/icVision/

Figure 4: The icVision Architecture: The core module (bot-

tom) is mainly for housekeeping and accessing & distribut-

ing the robot’s camera images and motor positions. Object

detection is performed in the ﬁlter modules (top), by seg-

menting the object of interest from the background. A typ-

ical work ﬂow is shown (right): input images are retrieved

from the cameras, the speciﬁc object is detected by a trained

Filter Module, before the outputs (together with the robot’s

current pose) are used to estimate a 3D position using a lo-

calisation module. The object is then placed in the world

model.

using our Cartesian Genetic Programming (CGP) im-

plementation called CGP-IP, in combination with the

OpenCV (Bradski, 2000) library, to detect one spe-

ciﬁc object in a variety of real-life situations (see Fig-

ure 2) – hence also identifying it. The framework can

run multiple ﬁlter modules in parallel. A variety of

ﬁlters have been learnt and most of them are able to

perform the object detection in near real-time.

To interact with the objects the robot also needs to

know where the object is located. Developing an ap-

proach to perform robust localisation to be deployed

on a real humanoid robot is necessary to provide the

necessary inputs for on-line motion planning, reach-

ing, and object manipulation. icVision provides mod-

ules to estimate the 3D position based on the robot’s

pose and the location of object in the camera images.

3.3 Integration

The sensory and motor sides establish quite a few

capabilities by themselves, yet to grasp objects suc-

cessfully while avoiding obstacles they need to work

closely together. The continuous tracking of obstacles

and the target object is required to create a reactive

reaching behaviour which adapts in real-time to the

changes of the environment.

We created interfaces between MoBeE and icVi-

sion allowing for a continuous visual based locali-

sation of the detected objects to be propagated into

the world model. This basic eye-hand coordination

allows for an adaptation while executing the reach-

ing behaviour to changing circumstances, improving

our robot’s autonomy. Figure 5 gives an overview of

ReactiveReachingandGraspingonaHumanoid-TowardsClosingtheAction-PerceptionLoopontheiCub

105

the research in our lab. Various modules (grey boxes)

were developed and integrated into a functional sys-

tem. The perception side consists mainly of object de-

tection and localisation (top area, depicted in green).

The bottom area (yellow) shows the action/motion

side, while the blue area on the right represents the

memory modules, including a world model. The var-

ious modules are:

• Object Models, Detection and Identiﬁcation: as

mentioned above, the detection and identiﬁcation

of objects is a hard problem. To perform these

tasks CGP-IP (Cartesian Genetic Programming

for Image Processing) (Harding et al., 2013) is

used. It provides a machine learning approach to

building visual object models, which can be con-

verted into executable code, in both supervised

and unsupervised fashion (Leitner et al., 2012a).

The resulting program performs the segmentation

of the camera images for a speciﬁc object.

• Object Localisation: by using the image coordi-

nates of the detected object from the two cameras

together with the current robot’s pose, the position

of the object can be estimated in Cartesian space

wrt. the robot’s reference frame. Instead of a cal-

ibration for each single camera, the stereo system

and the kinematic chain, we incorporate a module

that learns to predict these from a training set. Our

system has been shown to estimate these positions

with a technique based on genetic programming

(Leitner et al., 2012d) and an artiﬁcial neural net-

work estimators (Leitner et al., 2012c). After the

object is detected in the camera images the loca-

tion of an object is estimated and the world model

is updated.

• Action Repertoire: herein we only use a light-

weight, easy-to-use, one-shot grasping system

(LEOGrasper

), which has been used extensively

at IDSIA (Figure 6). It can be conﬁgured to per-

Source code available at: https://github.com/Juxi/iCub/

Figure 5: Overview of the modules and the on-going re-

search towards a functional eye-hand coordination system

on the iCub. The modules are shown by subsystem: per-

ception (green), action (yellow) and memory (blue).

form a variety of grasps, all requiring to close the

ﬁngers in a coordinated fashion. The iCub incor-

porates touch sensors on the ﬁngertips, due to the

high noise, we use the error reported by the PID

controllers of the ﬁnger motors to know when they

are in contact with the object.

• World Model, Collision Avoidance and Motion

Generation: the world model keeps track of the

robot’s pose in space and the objects it has visu-

ally detected. Figure 7 shows this model includ-

ing the robot’s pose the static table, and the two

objects localised from vision. MoBeE is used to

safeguard the robot from (self-)collisions. It fur-

thermore allows to generate motion by forcing the

hand in operational space.

4 THE ROBOT

Our experimental platform is an open-hardware,

open-software humanoid robotic system developed

within various EU projects, called iCub (Tsagarakis

et al., 2007) (shown in Figure 1). It consists of two

arms and a head attached to a torso roughly the size

of a human child. The head and arms follow an an-

thropomorphic design and provide a high DOF sys-

tem that was designed to investigate human-like ob-

ject manipulation. It provides also a tool to investigate

human cognitive and sensorimotor development. To

allow for safe and ‘intelligent’ behaviours the robot’s

movements need to be coordinated closely with feed-

back from visual, tactile, and acoustic

sensors. The

iCub is an excellent experimental platform for cogni-

tive, developmental robotics and embodied artiﬁcial

intelligence (Metta et al., 2010). Due to its com-

plexity it is also a useful platform to test the scala-

bility and efﬁciency of various machine learning ap-

proaches when used together with real-life robotic

systems.

Acoustic feedback might be used when manipulation

objects to know if a grasp was successful.

Figure 6: Grasping a variety of objects successfully, such

as, tin cans, plastic cups and tea boxes. The module works

for both the right and left hand.

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

106

5 EXPERIMENTS & RESULTS

The inverse kinematics problem, i.e. placing the hand

at a given coordinate in operational space, can be

performed with previously available software on the

iCub, such as, the existing operational space con-

troller (Pattacini, 2011) or a roadmap-based approach

(Stollenga et al., 2013). These systems though re-

quire a very accurate calibration the mechanical sys-

tem, however some mechanical non linearities still

cannot be taken into account. In addition they are not

yet coupled to visual localisation, meaning there is no

easy way to adapt the reach during execution.

The ﬁrst experiment shows that with the herein

presented system we are able to reactively move the

arm out of harms way when the environment changes.

We then show how we can use this system to reac-

tively reach and grasp these objects when we change

their type from ‘obstacle’ to ‘target’, therefore chang-

ing the ﬁctional forces calculated.

5.1 Avoiding a Moving Obstacle

Static objects in the environment can be added di-

rectly into MoBeE’s world model. Once, e.g. the

table, is in the model, actions and behaviours are

adapted due to computed constraint forces. These

forces, f

(t) in (1), which repel the robot from col-

lisions with the table, governs the actual movement

of the robot. This way we are able to send arbitrary

motions to our system, while ensuring the safety of

our robot (this has recently been shown to provide a

good reinforcement signal for learning robot reaching

behaviours (Pathak et al., 2013; Frank et al., 2014)).

The presented system has the same functionality also

for arbitrary, non-static objects. After detection in

both cameras the object’s location is estimated (icVi-

sion) and propagated to MoBeE. The ﬁctional forces

Figure 7: Showing the visual output of the MoBeE world

model during one of our experiments. Parts in red indi-

cate (an impeding) collision with the environment (or it-

self). The inset shows the actual scene. (See video: https://

www.youtube.com/watch?v=w qDH5tSe7g).

Figure 8: The reactive control of the left arm, permit-

ting the iCub to stay clear of the non-static ‘obstacle’, as

well as the table. (See video: https://www.youtube.com/

watch?v=w qDH5tSe7g).

are calculated to avoid impeding collisions. Figure 7

shows how the localised object is in the way of the

arm and the hand. To ensure the safety of the rather

fragile ﬁngers, a sphere around the end-effector can

be seen. It is red, indicating a possible collision, be-

cause the sphere intersects with the object. The same

is valid for the lower arm. The forces, calculated at

each body part using Jacobians, push the intersecting

geometries away from each other, leading to a forc-

ing of the hand (and arm) away from the obstacle.

Figure 8 shows how the the robot’s arm is avoiding

a non-stationary obstacle. The arm is ‘pushed’ aside

at the beginning, when the cup is moved close to the

arm. It does so until the arm reaches its limit, then the

forces cumulate and the end-effector is ‘forced’ up-

wards to continue avoiding the obstacle. Without an

obstacle the arm starts to settle back into its resting

pose q

∗

5.2 Reaching and Grasping Objects

This next experiment is on a simple reactive pick-and-

place routine for the iCub. Similarly to the above ex-

periment we are using MoBeE to adapt the reaching

behaviour while the object is moved. To do this we

change the type of the object within the world model

from ‘obstacle’ into ‘target’. Due to this change there

is no repelling force calculated between the object and

the robot parts. In fact we can now use the vector

from the end-effector to the target object as a force

that drives the hand towards a good grasping position.

MoBeE also allows to trigger certain responses

when collisions occur. In the case, when we want

the robot to pick-up the object, we can active a grasp

subsystem whenever the hand is in the close vicin-

ity of the object. We are using a prototypical power

grasp style hand-closing action, which has been used

successfully in various demos and videos.

Figure 6

shows the iCub successfully picking up (by adding an

extra upwards force) various objects using our grasp-

ing subsystem, executing the same action.

See videos at: http://robotics.idsia.ch/media/

ReactiveReachingandGraspingonaHumanoid-TowardsClosingtheAction-PerceptionLoopontheiCub

107

Our robot frameworks are able to track multiple

objects at the same time, which is also visible in Fig-

ure 7, where it tracks both the cup and the tea box. By

simply changing the type of the object within MoBeE

the robot reaches for a certain object while avoiding

the other.

6 CONCLUSIONS

Herein we present our on-going research towards vi-

sual guided object manipulation with the iCub. A

tightly integrated sensorimotor system, based on two

frameworks developed over the past years, enables the

robot to perform a simple pick-and-place task. The

robot reaches to detected objects, placed at random

positions on a table.

Our implementation enables the robot to adapt to

changes in the environment. Through this it safe-

guards the iCub from unwanted interactions – i.e. col-

lisions. We do this by integrating the visual system

with the motor side by using an attractor dynamic

based on the robot’s pose and a model of the world.

This way we achieve a level of eye-hand coordination

not previously seen on the iCub.

In the future we would like to integrate some ma-

chine learning to further improve the object manipula-

tion skills of our robotic system. Improving the pred-

ication and selection of actions will lead to a more

adaptive, versatile robot. Furthermore it might be

of interest to investigate an even tighter sensorimo-

tor coupling, e.g. avoiding translation into operational

space by working in vision/conﬁguration space.

ACKNOWLEDGEMENTS

The authors would like to thank the European Com-

mission for supporting this research under grant no.

FP7-IST-IP-231722, ‘Intrinsically Motivated Cumu-

lative Learning Versatile Robots’ (IM-CLeVeR).

REFERENCES

Ambrose, R., Wilcox, B., Reed, B., Matthies, L., Lavery,

D., and Korsmeyer, D. (2012). Robotics, tele-robotics,

and autonomous systems roadmap. Technical report,

NASA.

Bay, H., Tuytelaars, T., and Van Gool, L. (2006). Surf:

Speeded up robust features. Computer Vision – ECCV

2006, pages 404–417.

Bradski, G. (2000). The OpenCV Library. Dr. Dobb’s Jour-

nal of Software Tools.

Brooks, R. (1991). Intelligence without representation. Ar-

tiﬁcial intelligence, 47(1):139–159.

Carbone, G. (2013). Grasping in robotics, volume 10 of

Mechanisms and Machine Science. Springer.

Chaumette, F. and Hutchinson, S. (2006). Visual servo con-

trol, part i: Basic approaches. IEEE Robotics & Au-

tomation Magazine, 13(4):82–90.

Cipolla, R., Battiato, S., and Farinella, G. M. (2010). Com-

puter Vision: Detection, Recognition and Reconstruc-

tion, volume 285. Springer.

Corke, P. (2011). Robotics, Vision and Control, volume 73

of Springer Tracts in Advanced Robotics. Springer.

Davison, A. J. and Murray, D. W. (2002). Simultaneous lo-

calization and map-building using active vision. IEEE

Trans. on Pattern Analysis and Machine Intelligence,

24(7):865–880.

De Santis, A., Albu-Schaffer, A., Ott, C., Siciliano, B., and

Hirzinger, G. (2007). The skeleton algorithm for self-

collision avoidance of a humanoid manipulator. In

Proc. of the IEEE/ASME International Conferenced

on Advanced Intelligent Mechatronics.

Dietrich, A., Wimbock, T., Taubig, H., Albu-Schaffer, A.,

and Hirzinger, G. (2011). Extensions to reactive self-

collision avoidance for torque and position controlled

humanoids. In Proc. of the IEEE International Con-

ference on Robotics and Automation (ICRA).

Forssberg, H., Eliasson, A., Kinoshita, H., Johansson, R.,

and Westling, G. (1991). Development of human pre-

cision grip i: basic coordination of force. Experimen-

tal Brain Research, 85(2):451.

Frank, M., Leitner, J., Stollenga, M., Frster, A., and

Schmidhuber, J. (2014). Curiosity driven reinforce-

ment learning for motion planning on humanoids.

Frontiers in Neurorobotics, 7(25).

Frank, M., Leitner, J., Stollenga, M., Harding, S., F

orster,

A., and Schmidhuber, J. (2012). The modular be-

havioral environment for humanoids and other robots

(MoBeE). In Proc. of the Int’l. Conf. on Informatics

in Control, Automation & Robotics (ICINCO).

Harding, S., Leitner, J., and Schmidhuber, J. (2013). Carte-

sian genetic programming for image processing. In

Riolo, R., Vladislavleva, E., Ritchie, M. D., and

Moore, J. H., editors, Genetic Programming Theory

and Practice X, Genetic and Evolutionary Computa-

tion, pages 31–44. Springer New York.

Hartley, R. and Zisserman, A. (2000). Multiple view geom-

etry in computer vision. Cambridge Univ Press.

Hudson, N., Howard, T., Ma, J., Jain, A., Bajracharya, M.,

Myint, S., Kuo, C., Matthies, L., Backes, P., Hebert,

P., Fuchs, T., and Burdick, J. (2012). End-to-end dex-

terous manipulation with deliberate interactive esti-

mation. In Proc. of the IEEE International Confer-

ence on Robotics and Automation (ICRA).

Karlsson, N., Di Bernardo, E., Ostrowski, J., Goncalves,

L., Pirjanian, P., and Munich, M. (2005). The vS-

LAM Algorithm for Robust Localization and Map-

ping. In Proc. of the IEEE International Conference

on Robotics and Automation (ICRA).

Kemp, C., Edsinger, A., and Torres-Jara, E. (2007). Chal-

lenges for robot manipulation in human environments

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

108

[grand challenges of robotics]. IEEE Robotics & Au-

tomation Magazine, 14(1).

Khatib, O. (1986). Real-time obstacle avoidance for manip-

ulators and mobile robots. The international journal

of robotics research, 5(1):90.

Kragic, D. and Vincze, M. (2009). Vision for robotics.

Foundations and Trends in Robotics, 1(1):1–78.

LaValle, S. (2006). Planning algorithms. Cambridge Univ

Pr.

Leitner, J., Chandrashekhariah, P., Harding, S., Frank, M.,

Spina, G., Forster, A., Triesch, J., and Schmidhuber, J.

(2012a). Autonomous learning of robust visual object

detection and identiﬁcation on a humanoid. In Proc.

of the IEEE International Conference on Development

and Learning and Epigenetic Robotics (ICDL).

Leitner, J., Harding, S., Frank, M., F

orster, A., and Schmid-

huber, J. (2012b). icVision: A Modular Vision System

for Cognitive Robotics Research. In Proc. of the Inter-

national Conference on Cognitive Systems (CogSys).

Leitner, J., Harding, S., Frank, M., F

orster, A., and Schmid-

huber, J. (2012c). Learning spatial object localisation

from vision on a humanoid robot. International Jour-

nal of Advanced Robotic Systems, 9.

Leitner, J., Harding, S., Frank, M., F

orster, A., and Schmid-

huber, J. (2012d). Transferring spatial perception be-

tween robots operating in a shared workspace. In

Proc. of the IEEE/RSJ International Conference on

Intelligent Robots and Systems (IROS).

Lowe, D. (1999). Object Recognition from Local Scale-

Invariant Features. In Proc. of the International Con-

ference on Computer Vision.

Maitin-Shepard, J., Cusumano-Towner, M., Lei, J., and

Abbeel, P. (2010). Cloth grasp point detection based

on multiple-view geometric cues with application to

robotic towel folding. In Proc. of the IEEE In-

ternational Conference on Robotics and Automation

(ICRA).

Metta, G., Fitzpatrick, P., and Natale, L. (2006). YARP:

Yet Another Robot Platform. International Journal of

Advanced Robotic Systems, 3(1).

Metta, G., Natale, L., Nori, F., Sandini, G., Vernon, D.,

Fadiga, L., von Hofsten, C., Rosander, K., Lopes, M.,

Santos-Victor, J., Bernardino, A., and Montesano, L.

(2010). The iCub humanoid robot: An open-systems

platform for research in cognitive development. Neu-

ral Networks, 23(8-9):1125–1134.

Panin, G., Ladikos, A., and Knoll, A. (2006). An efﬁcient

and robust real-time contour tracking system. In Proc.

of the IEEE International Conference on Computer

Vision Systems.

Pathak, S., Pulina, L., Metta, G., and Tacchella, A. (2013).

Ensuring safety of policies learned by reinforcement:

Reaching objects in the presence of obstacles with the

icub. In Proc. of the IEEE/RSJ International Confer-

ence on Intelligent Robots and Systems (IROS).

Pattacini, U. (2011). Modular Cartesian Controllers for

Humanoid Robots: Design and Implementation on the

iCub. PhD thesis, Italian Institute of Technology, Gen-

ova.

Plumert, J. and Spencer, J. (2007). The emerging spatial

mind. Oxford University Press, USA.

Saxena, A., Driemeyer, J., and Ng, A. (2008). Robotic

grasping of novel objects using vision. The In-

ternational Journal of Robotics Research (IJRR),

27(2):157.

Schoner, G. and Dose, M. (1992). A dynamical systems

approach to task-level system integration used to plan

and control autonomous vehicle motion. Robotics and

Autonomous Systems, 10(4):253–267.

Stollenga, M., Pape, L., Frank, M., Leitner, J., F

orster, A.,

and Schmidhuber, J. (2013). Task-relevant roadmaps:

A framework for humanoid motion planning. In Proc.

of the IEEE/RSJ International Conference on Intelli-

gent Robots and Systems (IROS).

uckler, J., Badami, I., Droeschel, D., Gr

ave, K., Holz,

D., McElhone, M., Nieuwenhuisen, M., Schreiber,

M., Schwarz, M., and Behnke, S. (2013). Nim-

bro@home: Winning team of the robocup@home

competition 2012. In Robot Soccer World Cup XVI,

pages 94–105. Springer.

Tsagarakis, N., Metta, G., Sandini, G., Vernon, D., Beira,

R., Becchi, F., Righetti, L., Santos-Victor, J., Ijspeert,

A., Carrozza, M., and Caldwell, D. (2007). iCub: the

design and realization of an open humanoid platform

for cognitive and neuroscience research. Advanced

Robotics, 21:1151–1175.

Vahrenkamp, N., Wieland, S., Azad, P., Gonzalez, D., As-

four, T., and Dillmann, R. (2008). Visual servoing for

humanoid grasping and manipulation tasks. In Proc.

of the IEEE International Conference on Humanoid

Robots, pages 406–412.

ReactiveReachingandGraspingonaHumanoid-TowardsClosingtheAction-PerceptionLoopontheiCub

109