Supplementary material for:
Dynamical self-consistency leads to behavioral development and emergent social interactions in robots.

Ralf Der and Georg Martius

The paper will appear in Proceedings of the 6th Joint IEEE International Conference on Development and Learning and on Epigenetic Robotics. A personal copy of the paper can by found here.

Abstract: We present an approach that enables robots to self-organize their sensorimotor behavior from scratch without providing specific information about neither the robot nor its environment. This is achieved by a simple neural control law that increases the consistency between external sensor dynamics and internal neural dynamics of the utterly simple controller. In this way, the embodiment and the agent-environment coupling are the only source of individual development. We show how an anthropomorphic tendon driven arm-shoulder system develops different behaviors depending on that coupling. For instance: given a bottle half-filled with water, the arm starts to shake it, driven by the physical response of the water. When attaching a brush, the arm can be manipulated into wiping a table, and when connected to a revolvable wheel it finds out how to rotate it. Thus, the robot may be said to discover the affordances of the world. When allowing two (simulated) humanoid robots to interact physically, they engage into a joint behavior development leading to, for instance, spontaneous cooperation. More social effects are observed if the robots can visually perceive each other. Although, as an observer, it is tempting to attribute an apparent intentionality, there is nothing of the kind put in. As a conclusion, we argue that emergent behavior may be much less rooted in explicit intentions, internal motivations, or specific reward systems than is commonly believed.

1  Overview

Overview 29Compiled clip of all arm-shoulder experimentsVideo 1
Bottle swingExcitation of a circular pendulum modeVideo 2
Shaking verticallyA half filled bottle is vertically attached to the tip of the arm: shaking of the bottle mainly along its axisVideo 3
Shaking horizontallySame as above but with horizontal attachmentVideo 4
Rotating wheelArm attached to a revolvable bar/wheelVideo 5
Rotating wheel IIParallel wheel-arm arrangementVideo 6
Rotating wheel IIIDifferent rotation frequenciesVideo 7
Wiping tableArm with brush starts to wipe a tableVideo 8
Wiping table modesDifferent wiping patterns from reloaded controllersVideo 9
FreeNo external forces applied: pseudo-random sequences of reaching-type behaviorVideo 10
Crawling humanoidHumanoid robot on the ground develops a crawling behavior from scratchVideo 11
Humanoids at a wheelTwo humanoid robots hold on to the cranks of a wheel and jointly rotate itVideo 12
Socializing IHarmony in emerging behavior of two humanoids suspended on elastic ropesVideo 13
Socializing IIEmerging patters with inverted visionVideo 14
Socializing IIIOn stools, one robot is weakened and gets perturbed repeatedlyVideo 15
Socializing IVSame as above, but with delayed visionVideo 16
Alien body effectsTwo humanoids percieve only sensors of other robot (inversed sign)Video 17
FightersTwo humanoids fightingVideo 18

2  Videos

Video 1: Overview video summarizing the experimental results on the Myorobotic Arm of the paper. This video provides a demonstration of the control paradigm applied to the tendon driven anthropomorphic arm. Compliant control self-organizes from a dynamical interplay between robot, environment and neural plasticity of the controller. [mp4 video file]
back to List of Videos

Video 2: Bottle swing: Excitation of a circular pendulum mode. The suspended bottle, once exited to swing a little bit, exerts forces onto the arm, which are proagating into the controller through the plasticity rule. This leads eventually to a coherent swinging motion. [mp4 video file]
back to List of Videos

Video 3: Shaking vertically: A half filled bottle is vertically attached to the tip of the arm. Emergent shaking behavior of the bottle mainly along its axis. [mp4 video file]
back to List of Videos

Video 4: Shaking horizontally: Same as above but with horizontal attachment. The main shaking direction is now horizontal. Apart from the water in the bottle, the physical conditions are the same (weight of the bottle) but the emerging motion patterns are seen to be different from the horizontal scenario, so as if the robot feels the specific reactions of the the water to the motions of the arm. Again, this demonstrates how the physical reactions of the environment (the inertia forces when the water hits the walls of the bottle) enter the learning mechanism. [mp4 video file]
back to List of Videos

Video 5: Rotating wheel: Arm is attached to a revolvable bar/wheel. Initially the connection between the arm and the wheel was rather loose so that for small movements there is no physical reaction from the rotation of the wheel. After improving this connection, an initial push by the experimenter was sufficient for excite a stable rotation mode. [mp4 video file]
back to List of Videos

Video 6: Rotating wheel II: Arm and wheel are now arranged in a parallel setup. After an initial push, a stable rotation behavior is observed. Later it is demonstrated that the system can be switched between forward and backward rotation mode. This is a direct consequence of the time inversion symmetry of the underlying physics of the mechanical system, i. e. the arm connected physically to the wheel. [mp4 video file]
back to List of Videos

Video 7: Rotating wheel III: Different rotation frequencies By changing the time-delay d of the delay-sensors, the frequency of the emerging cyclic motion can be adjusted, resulting in different rotation velocities. [mp4 video file]
back to List of Videos

Video 8: Wiping table: The arm with a brush starts to wipe a table. Through the combination of the manual force and the constraints given by table surface, the robot is driven into the two-dimensional wiping mode. Later in this video the robot is guided into a different behavior, which persists for a short time before it switches to a new behavior. [mp4 video file]
back to List of Videos

Video 9: Wiping table modes: Different wiping patterns from reloaded controllers. During earlier runs we took snapshots of the controller weights C when a new wiping mode occured. In this video we show that by simply reloaded the controllers (and keeping their weights fixed) the robot transitions smoothly into the respective attractor behavior. [mp4 video file]
back to List of Videos

Video 10: Free: without external forces a pseudo-random sequences of reaching-type behavior may be observed. The emerging behaviors are a combination from the physical with the learning dynamics. Without any dynamically structured response form the physics, the system goes into a kind of searching motion. [mp4 video file]
back to List of Videos

Video 11: The humanoid robot on the floor developing a crawling behavior. During the initial period it is seen how, from the initialization condition, small movements get amplified into a coherent movement. These get increasingly shaped to fit the situation. From time 4:30 on a stable crawling motions is observed. [mp4 video file]
back to List of Videos

Video 12: Humanoids at a wheel: Emerging communications by force exchange. Both robots are connected to the wheel, each to one of the cranks. The robots have no information about their partner. Yet they manage to cooperate by “feeling” the others reactive forces. This even works if the robots are not supported by the stool (B) (muscle forces doubled). [mp4 video file (a)][mp4 video file (b)]
back to List of Videos

Video 13: Socializing I: Harmony in emerging behavior of two humanoids suspended on elastic ropes Now the robots have vision meaning that they receive the joint angle values of their mates as additional sensor values. Behavior is now built essentially by watching each other. Depending on the parameters, many highly coordinated behaviors are observed. At time 07:40, perturbations by external forces (the red and green dots) were repeatedly applied for demonstrating the recovery to new coordinated motion patterns. In particular, from time 08:00 on, the recovery to a new mode is observed after very heavy perturbations, even though the robots are perturbing each other. [mp4 video file]
back to List of Videos

Video 14: Socializing II: Emerging patters with inverted vision The guiding influence of the vision sensors even persists if the sign of the vision signal is inverted. [mp4 video file]
back to List of Videos

Video 15: Socializing III: On stools, one robot is weakened and gets perturbed repeatedly Demonstrating again the spontaneous correlation of motions by the influence of the vision channel. [mp4 video file]
back to List of Videos

Video 16: Socializing IV: Same as Video 15, but with delayed vision. With a delay in the vision signal, convergence toward a coherent limit cycle behavior is often observed. [mp4 video file]
back to List of Videos

Video 17: Alien body effects: Two humanoids, having as sensor values only the joint angles of its opponent (with inversed sign). Each robot has to realize the learning and control of its body exclusively on the basis of its mate’s movements, as reported by the visual system. In a sense, each robot is building its behavior on the illusion that the body of the other one is its own. [mp4 video file]
back to List of Videos

Video 18: Fighters: Two humanoids in tight interaction. Each robot carries a magnet at its fists which connects upon contact with the iron trunk of the opponent. Then, the magnet is active for 20 seconds, remaining inactive for another 30 seconds. Each robot is driven by the one-layer neurocontroller with the generic learning rule, the same as in the wheel rotation example. Different from the latter, there is no convergence toward a stable periodic behavior, as there is no dynamical structuring from the reaction of the environment. So, the emerging motion patterns are essentially long lived transients between potential limit cycle attractors. [mp4 video file]
back to List of Videos

This document was translated from LATEX by HEVEA.