Contents 
See also 
Look here if your video player has troubles in playing the videos files.Check also the video page for our book with many videos (all with flash player). 
To find how autonomous behavior and its development can be realized by a brain or an artificial neural network is a fundamental challenge for neuroscience and robotics. Commonly, the selforganized unfolding of behavior is explained by postulating special concepts like internal drives, curiosity, specific reward systems, or selective pressures. We propose a simple, local, and biologically plausible synaptic mechanism that enables an embodied agent to selforganize its individual sensorimotor development without recourse to such higher level constructs. When applied to robotic systems, a rich spectrum of rhythmic behaviors emerges, ranging from locomotion patterns to the spontaneous cooperation between partners. A possible utilization of our novel mechanism in nature would lead to a new understanding of the early stages of sensorimotor development.
See also Videos page of the paper! 
Information theory is a powerful tool to express principles to drive autonomous systems because it is domain invariant and allows for an intuitive interpretation. This paper studies the use of the predictive information (PI), also called excess entropy or effective measure complexity, of the sensorimotor process as a driving force to generate behavior. We study nonlinear and nonstationary systems and introduce the timelocal predicting information (TiPI) which allows us to derive exact results together with explicit update rules for the parameters of the controller in the dynamical systems framework. In this way the information principle, formulated at the level of behavior, is translated to the dynamics of the synapses. We underpin our results with a number of case studies with highdimensional robotic systems. We show the spontaneous cooperativity in a complex physical system with decentralized control. Moreover, a jointly controlled humanoid robot develops a high behavioral variety depending on its physics and the environment it is dynamically embedded into. The behavior can be decomposed into a succession of lowdimensional modes that increasingly explore the behavior space. This is a promising way to avoid the curse of dimensionality which hinders learning systems to scale well.
Humanoid in different environemnts MPEG4H264 (15.1 MB) © Georg Martius 
See also Videos page of the paper! 
Locomotion 1 DIVX4 (13.3 MB) © Georg Martius 
Locomotion 1 DIVX4 (12.4 MB) © Georg Martius 
Turning DIVX4 (27.2 MB) © Georg Martius 
See also Videos page of the paper! Here are some things we have observed just after connecting our algorithm. Note that the robot does not move when grasped because the acceleration sensors (are only sensor information) stay more or less constant. 
Compilation DIVX4 (10.8 MB) © Georg Martius 
With weight DIVX4 (9.2 MB) © Georg Martius 
We had one day to connect our controller to the robot and observe what happens. Note how sensitive the robot it its acceleration sensors when touched or moved externally. See also the video page for our book. 
Back Flip AVI (0.5 MB) (0.1 MB) DIVX4 (0.3 MB) © Ralf Der 
Rolling over AVI (0.6 MB) (0.1 MB) DIVX4 (0.4 MB) © Ralf Der 
In full action AVI (5.2 MB) (1.0 MB) DIVX4 (3.3 MB) © Ralf Der 
Motion patterns emerging in the course of time if the robot is left alone in free space. The robot is controlled by our selfregulating neural network according to the homeokinesis principle. The joints are actuated by simulated servomotors. Sensor values are the joint angles. There is no other information available to the robot. 
MPEG2 (124.2 MB) DIVX4 (5.9 MB) © Ralf Der 
DIVX4 (3.3 MB) © Ralf Der 
AVI (7.5 MB) (1.6 MB) DIVX4 (5.0 MB) © Ralf Der 
AVI (6.1 MB) (1.3 MB) DIVX4 (4.0 MB) © Ralf Der 
A selfrescue scenario: AVI (1.7 MB) (0.3 MB) DIVX4 (1.0 MB) © Ralf Der 
The robot in a narrow pit develops after some time motion patterns which may help it to get out of the impasse.

A bar excercise scene AVI (13.0 MB) (2.4 MB) DIVX4 (8.6 MB) (2.2 MB) © Ralf Der 
Two hours later AVI (5.0 MB) (0.9 MB) DIVX4 (3.2 MB) © Ralf Der 
Dog fixed in the air AVI (1.3 MB) (0.2 MB) MPEG2 (1.5 MB) DIVX4 (0.8 MB) © Georg Martius 
Dog fixed in the air II AVI (1.1 MB) (0.2 MB) MPEG2 (1.4 MB) DIVX4 (0.7 MB) © Georg Martius 
Dog surmounting a barrier 1 AVI (6.7 MB) (1.3 MB) DIVX4 (8.8 MB) © Ralf Der & Georg Martius 
Dog surmounting a barrier 2 AVI (8.0 MB) (1.4 MB) MPEG2 (10.7 MB) DIVX4 (5.4 MB) © Ralf Der & Georg Martius 
The hippodog AVI (7.6 MB) (1.5 MB) DIVX4 (10.0 MB) © Ralf Der & Georg Martius 
AVI (9.9 MB) (1.8 MB) MPEG1 (18.7 MB) © Ralf Der 
Rolling Wheelie AVI (3.2 MB) (0.7 MB) MPEG2 (4.1 MB) DIVX4 (2.1 MB) © Georg Martius 
AVI (0.0 MB) (0.0 MB) MPEG2 (5.7 MB) © Ralf Der & Frank Hesse 
Simulation of a human hand with multiple degrees of freedom (in this simulation only 6 degrees of freedom are used). The hand is equipped with motion sensors at all joints. The infrared sensors at the finger tips are not used in this experiment. It is operated in a fully exploratory mode with or without a manipulated object. 
The subcritical behaviour: Low PI. Incoherent fluctuations. Low activity (narrow range of sensor values) and some predictability C[0,0]=C[1,1]=0.8 and C[0,1]=C[1,0]=0.0 AVI (2.0 MB) (0.4 MB) DIVX4 (1.3 MB) © Frank Güttler & Ralf Der 
Near optimal behaviour: High PI due to a good balance between activity (wide range of sensor values) and predictability. C[0,0]=C[1,1]=1.0 and C[0,1]=C[1,0]=0.07 DIVX4 (2.5 MB) © Frank Güttler & Ralf Der 
The supracritical behaviour: Low PI due to high activity but only switching of sensor values and reduced predictability System is caught in a behavioral mode. C[0,0]=C[1,1]=1.4 and C[0,1]=C[1,0]=0.0 DIVX4 (1.2 MB) © Frank Güttler & Ralf Der 
The subcritical behaviour: C[0,0]=C[1,1]=0.8 and C[0,1]=C[1,0]=0.0 MJPEG (8.0 MB) DIVX4 (0.4 MB) (0.1 MB) © Frank Güttler & Ralf Der DIVXFormat (320x240, 0,3 MB)  The critical behaviour: C[0,0]=C[1,1]=1.0 and C[0,1]=C[1,0]=0.1 AVI (1.8 MB) (0.5 MB) DIVX4 (1.1 MB) (0.3 MB) © Frank Güttler & Ralf Der 
The supracritical behaviour: C[0,0]=C[1,1]=1.3 and C[0,1]=C[1,0]=0.2 AVI (2.5 MB) (0.6 MB) DIVX4 (1.6 MB) (0.4 MB) © Frank Güttler & Ralf Der 
Demo 1 AVI (3.0 MB) (0.6 MB) MPEG2 (3.8 MB) DIVX4 (1.9 MB) © Frank Güttler & Ralf Der 
Demo 2 AVI (3.0 MB) (0.6 MB) MPEG2 (3.8 MB) DIVX4 (1.9 MB) © Frank Güttler & Ralf Der 
Demo 3 AVI (2.7 MB) (0.5 MB) MPEG2 (3.5 MB) DIVX4 (1.8 MB) © Frank Güttler & Ralf Der 
Life is starting in the zoo AVI (23.7 MB) MPEG2 (28.4 MB) © Georg Martius 
The zoo in full action AVI (12.6 MB) MPEG2 (13.1 MB) © Georg Martius 
The somewhat jerky run of the video streams originates from technical problems when recording it. The original computer simulation is free of these artefacts.
AVI (3.8 MB) (0.8 MB) MPEG1 (7.0 MB) © Georg Martius 
AVI (7.0 MB) (1.6 MB) MPEG1 (13.0 MB) © def 
AVI (7.9 MB) (1.6 MB) MPEG1 (14.9 MB) DIVX4 (10.3 MB) © def 
We see that each of the agents develops a kind of behavior of its own which is emerging from the interplay of its body with the environment. Behaviors are largely modified by the interactions with the obstacles in the arena and with other agents. The latter kind of interactions makes our zoo a highly dynamic environment.
=> back to top
AVI (12.1 MB)
MPEG2 (13.3 MB)
© Frank Hesse
The parameters of the world model demonstrate how the model relearns in order to reflect the different reactions of the complex body (the string of gyros) to the applied forces. In the beginning the motions are very slow so that the response of the body is weak. This is reflected by the small values of the matrix elements of . When the snake is transiting into the rotational mode the parameters of both the controller and the world model drastically change, the latter reaching essentially a rotation matrix.
AVI (9.2 MB)
MPEG2 (9.6 MB)
DIVX4 (6.7 MB)
© Frank Hesse
All three snakes are probing into different frequencies, which are seen to be transient.
AVI (5.0 MB)
MPEG2 (5.3 MB)
© Frank Hesse
AVI (0.7 MB) (0.2 MB)
MPEG1 (1.3 MB)
© Georg Martius
AVI (14.9 MB) (2.8 MB)
MPEG1 (28.2 MB)
© Georg Martius
SVGA resolution (1024x768)
AVI (15.0 MB) (2.8 MB)
MPEG2 (41.5 MB)
© Georg Martius
AVI (13.9 MB) (2.5 MB)
MPEG2 (41.2 MB)
© Georg Martius
AVI (16.6 MB) (3.0 MB)
MPEG2 (48.9 MB)
DIVX4 (34.5 MB)
© Georg Martius
AVI (15.6 MB) (2.9 MB)
MPEG2 (46.0 MB)
DIVX4 (20.5 MB)
© Georg Martius
AVI (18.4 MB) (3.3 MB)
MPEG2 (24.1 MB)
DIVX4 (12.2 MB)
© Georg Martius
Behavior 1

Behavior 2

Demonstration sensor control 1

Demonstration sensor control 2

AVI (1.8 MB) © Ralf Der & Rene Liebscher 
AVI (5.5 MB) © Ralf Der & Rene Liebscher 
AVI (1.6 MB) © Ralf Der & Rene Liebscher 
In the videos you can see a number of different agents in an arena. Each of those is driven by a controller which is a neural network with a fast synaptic dynamics obtained from our general principle of homeokinesis realized as gradient descending an objective function which coarsely speaking is the weighted matrix norm
This completely domain invariant principle is seen to generate behaviors which are different for each agent and which change over time. The zoo may be run forever with the behaviors changing over time. The different actors of the scene are presented in their details and specific properties in the following below. Here we want to demonstrate that the paradigm can be translated into a feasible, extremely robust algorithm despite of the mathematical subtleties involved with , cf. Eq. 1, like the fact that it contains the inverse of the Jacobian matrix. Moreover a more detailed analysis of the dynamics shows that the crucial point of the concomitant learning of the world model and the controller is solved by the algorithm in a reliable and controlled way although the world is highly unstructured and dynamic.
=> back to topThe .avi videos (DivX Format) need an MPEG4Codec installed on your system like Xvid or DivX. Unfortunately, not all versions of the Windows Media Player have this codec by default, please install them to watch the videos. There are three suggestions:
It's necessary that you have administrator rights on your system (on linux plattforms it may differ). When using the KLite Codec Pack or the Free DivX software, you usually have to restart your system. The VLC media player plays immediately the videos after the installation process.
In order to playback the videos you need to have the codecs installed. Most distributions provide a codec package, sometimes called win32codecs. If you install them xine and mplayer can play the files. If you cannot find them go to the mplayer download page and download a Binary Codec package. If you have no root permissions you can store the codecs in ~/.mplayer/codecs.
Good luck!