Research Area "Neurocybernetics"

 
Home
Research topics
Projects
Tools
Publications
People
Links
Contact

Doctorate Programme
Institute of Cognitive Science
University of Osnabrück

Home - Downloads - Media - Contact - Links

NERD Toolkit

Neurodynamics and Evolutionary Robotics Development Toolkit



Closed-Chain Animat


This page provides supplemental material for a conference contribution of the SAB 2012:

Rempis, Pasemann.
"Evolving Variants of Neuro-Control using Constraint Masks"
Proceedings of the conference of simulation of adaptive behavior (SAB), August 2012.




Important Note:
To view the networks you need either the Network Editor of the NERD Toolkit (NERD Format *.onn) or a vector graphics programm (SVG) such as InkScape. The scaleable vector graphics images are comparably large and require a fast computer to be rendered in acceptable time! The initial network with the constraint mask can be downloaded for each experiment. This constraint mask requires the NERD Network Editor to be viewed!



This series of experiments used constraint masks to evolve different variants for a simple locomotion behavior. The animat had to overcome a hurdle track with obstacles of increasing difficulty. Due to the different constraint masks, much more variants of controllers have been found compared to a single series of experiments without constraints. A brief preview is shown in the animated gifs. The full behavior for the entire hurdle track can be viewed by clicking on the animated images.


Experiment M1: Reference Experiment with Minimal Constraint Mask (Configuration A)

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
M1
A
[ConstraintMask]
All Minimally constrained initial network with a single body type. A single master module is evolved, while all other modules are clones of that master module. All connections between the modules are forced to be rotation-symmetric.

closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller uses primarily the gyroscope and the force sensor. All other sensors can be disconnected without changing the behavior. This approach is the dominant solution in this experimental setting. Approximately 80 % of all evolved controllers used this approach.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller uses the gyroscope and a combination of the acceleration and the angular sensors. The other sensors can be disconnected without changing the behavior.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller uses the gyroscope and the angular sensor. All other sensors can be disconnected without changing the behavior.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller uses only the angular and the acceleration sensors.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller uses only the acceleration sensors. The force and angle sensors increase the efficiency of this approach, but the main behavior also works fine without them.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller is based on the angular sensors, assisted by the acceleration and gyrosope sensors that together significantly improve the basic angular driven behavior.

Experiment M2: Reference Experiment with Minimal Constraint Mask (Configuration B)

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
M2
B
[ConstraintMask]
All Minimally constrained initial network with three types of body parts. A single master module per type is evolved, while all other modules are clones of these master modules. All connections between the module-triplets are forced to be rotation-symmetric.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This controller relies on the gyroscope and the force sensors to create a organic looking forwards rolling.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This controller only requires a gyroscope sensor.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: The controller here also uses only the gyrosope sensors prodicing a different motion.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: Here, the gyroscopes and the force sensors are again used as successful solution, resulting in a star-like behavior.

Experiment V1: Communication Range of 1 Between Modules (Angular Sensor Only)

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V1
A
[ConstraintMask]
Angle Enforces the usage of neighboring groups and prevents the use of the own sensors. Additionally, the motor angle neuron is fixed and the behavior has to be realized using the motor torque neuron only.


Most controllers in this experiment behave quite similarly, but the implemented strategies in the networks differ from each other. This is an example where the real value of different controllers is not the number of different behaviors, but instead of different ways to do the same behavior. The resulting controllers tend to be autonomous pattern generators that only marginally react on their environment. A reason may be that to react on the environment the animat needs a way to get influenced by the environment. This is here only possible by forces that bend (partially) relaxed body segments and herewith change the angle sensors. In actuated joints, this is not possible because there the animat's forces keep the desired angles against disturbances from the environment and thus prevent a real interaction.

closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This controller implements an interesting mechanism to switch a behavior on and off. The neuron with the desired motor angle and the desired motor torque are directly coupled, so that both always have the same activity. This means that the joint can only be actively bent (both positive) or be passive without applying torque (both negative). In the latter state the joint can be stretched by the forces coming from its neightbors.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: Here, a sole torque control is realized, i.e. the angular sensor remains fixed at a certain activation and the motor is controlled by allowing more or less torque applied on the joint.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:

Experiment V2: Communication Range of 2 Between Modules (Angular Sensor Only)

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V2
A
[ConstraintMask]
Angle Allows only connections between every second neighbor, hereby defining two neuro-dynamically independent rings of groups.


In this experiment, each controller comprises two neurally-dynamically independent circuits. Body parts can only communicate with a range of 2, which means that neighboring body segments cannot send signals to each other. They "only" influence each other through the body. The interesting part of this experiment is to observe, how the two circuits synchronize and work together for a stable forwards locomotion.


closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
closed-chain-animat.avi closed-chain-animat.avi closed-chain-animat.avi Description: This network provides three coexisting variants of the behavior. An eight-shaped rolling, a fast elliptic rolling and a very slow, round rolling behavior. The different modes depend on the interplay of the two circuits. So, intermediate switches between the behaviors are possible due to interactions with the environment. The three videos here show all three behaviors. The first video also shows the transition of the circle-shaped motion to the eight-shaped motion (in the beginning of the video).
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This controller provides an organic-looking, slow motion.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: A kidney-shaped, contracting motion.

Experiment V3: Communication Range of 1 Between Modules (Acceleration Sensors Only)

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V3
A
[ConstraintMask]
Acceleration -


In these experiments, only acceleration sensors can be used to produce the behavior. Interestingly, the evolved controllers seem to be quite slow, but steady. A reason may be that slow robots have a more definite acceleration behavior of the body parts, compared to a fast, rapid moving animat, where the body parts are accelerated and decelerated abruptly. A controller based on acceleration sensors is more reactive to its environment, compared to the angular sensor driven controllers, but such controllers often have states where the behavior simply stopps due to missing changes in the acceleration sensors.


closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: Triangular-shaped locomotion. Due to the high center of gravity, it is difficult for this animat to overcome steep ramps.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: Interesting slow, but steady behavior. Does not fall over, but may get stuck if the acceleration sensors do not change any more (see end of video).

Experiment V4: Unevenly Sized Animat With Many Sensors (Direct Neightbor Communication)

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V4
B
[ConstraintMask]
All In addition to the angle sensor, each of the three differently sized segments uses only one of the other sensors (acceleration, gyroscope, force). So, each group of a triplet now has a different set of sensors.


This animat has segments of three different sizes. All segments of one size are equipped with similar sensor sets, either angle/acceleration, angle/oscilloscope or angle/force sensor. So, each sensor-pairing is only used on five of the fifteen body segments.


closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: A star-like locomotion behavior that overcomes even steep ramps by moving its center of gravity in a suitable way.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This controller makes the animat pile up on a single, small body segment and then to fall in the desired direction, hereby overcoming obstacles.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This little fellow literarily "catapults" itself forwards through the obstacle track. For this, most body parts are quite relaxed, which makes the animat look very thin and snake-like.


Experiment V5: Hybrid Animat Controlled With Only 5 Gyroscope Sensors

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V5
B
[ConstraintMask]
Gyroscope A single group of each triplet now has a gyroscope sensor. Accordingly, there are only 5 gyroscope sensor sets (x, y, z) to control all 15 segments. A special constraint ensures, that only networks evolve in which all motor neurons are influenced by (an arbitrarily long chain of) synapses coming from the gyroscope sensors.


In this configuration, the animat can only use five of the accelerometers (x, y, z), distributed equally on the animat's body. No other sensor is used. Here the difficulty is to use the few sensors to control also those body parts that are not using any own sensor. Constraints help to enforce that each sensor has a synaptic pathway between the acceleration sensors and all motors. This prevents the evaluation of controllers that do not have the required minimal neural structure to control all body parts.


closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
closed-chain-animat.avi Description: ...
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This controller uses "virtual limbs" to move forwards. The shape of the animat may be remindful of a horse.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description:

Experiment V6: Forced the Controller to Use a Neural Oscillator

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V6
A
[ConstraintMask]
Angle, Acceleration The master group is forced to use a neural oscillator with frequency control between the sensors and the motors. Therefore, the oscillators have to be connected suitably and a strategy for an effective synchronization has to be found.

closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
closed-chain-animat.avi Description: The neural oscillators are synchronized through the body and a synapse to the frequency neuron of the oscillator. After a short while, all oscillations are more or less in pairwise phases, both synchroneous joints belonging to opposite body segments of the animat.

Experiment V7: Using An Internal Neural Pulse for the Coordination

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V7
A
[ConstraintMask]
None The network is equipped with a structure that produces an activity pulse that passes through all groups. The frequency of the pulse can be evolved, as well as a strategy to use it to generate a locomotion.

closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
closed-chain-animat.avi Description: This controller uses the pulse to drive a single active bending center. The speed of the pulse is slowed to obtain a stable locomotion with only a few rollovers.

Experiment V8: Enforced Feed-forward Processing Column with Horizontal Excitation and Lateral Inhibition

Exp. Configuration Sensors Constraint Mask in Addition to the Default Mask
V8
A
[ConstraintMask]
Acceleration The master group is forced to develop an excitatory feed-forward structure between the sensors and the motors, whereas signals from the direct neighbors can influence that structure with inhibition only.

closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
closed-chain-animat.avi Description: Very interesting behavior. The animat tries to keep a circular shape, only interrupted by short elongations to speed up the robot. In this way, the robot can roll like a wheel. At obstacles, however, the animat elongates as far as needed to overcome the obstacle. Often, several attempts are needed, leading to run-ups and different approaches to overcome the obstacles. The activity patterns of this controller are also particularly interesting.
closed-chain-animat.avi
Network (NERD)
Network (SVG)
Search Space (SVG)
Description: This locomotion is comparably slow, but very steady. The hurdles are mastered mostly without backtracking or loopings.










Additional Variation Experiments

The following experiments have not been included in the publication, because the number of evolution runs, the experiment scenarios, the animat configuration or the evolution procedure differed too much from the described experiments above to allow a valid comparison. Nevertheless, these results additionally emphasizes the potential of the constraint mask approach to search explicitly and systematically for variants of neuro-control.

Some of these videos are in grayscale. This is only for technical reasons and has no deeper reason.


No Communication Connection Between Modules (Angular Sensor Only, no Hurdles)

closed-chain-animat.avi

closed-chain-animat.avi Note:

Larger animat with 20 links. Description to come...
closed-chain-animat.avi Note:

Larger animat with 20 links. Description to come...
closed-chain-animat.avi Note:

Larger animat with 20 links. Description to come...

Communication Range of 1 Between Modules (Angular Sensor Only, no Hurdles)

closed-chain-animat.avi
closed-chain-animat.avi
closed-chain-animat.avi
closed-chain-animat.avi

Communication Range of 1 with 20 Segments (Angular Sensor Only)

closed-chain-animat.avi Note:

Larger animat with 20 links. Description to come...

Communication Range of 1 AND 2 Between Modules (Angular Sensor Only)

closed-chain-animat.avi
closed-chain-animat.avi

Communication Range of 3 Between Modules (Angular Sensor Only)

closed-chain-animat.avi

Unevenly Sized Animat With All But the Angular Sensors (Direct Neightbor Communication)

This animat has segments of three different sizes. All segments of one size are equipped with similar sensor sets, either acceleration, oscilloscope or force sensor. The angular sensors have been removed.
closed-chain-animat.avi
closed-chain-animat.avi

Rotation-Symmetric Communication Between Modules (Acceleration Sensor Only)
closed-chain-animat.avi
closed-chain-animat.avi
closed-chain-animat.avi
closed-chain-animat.avi
closed-chain-animat.avi
closed-chain-animat.avi