Dennis Goldschmidt

Bericht Spezialisierungspraktikum Universität Göttingen, III. Physikalisches Institut

This article reports the work at the Third Institute of Physics at the University of Göttingen during bachelor’s specialization practicals which serve as a basis for preparation of the bachelor’s thesis. The article introduces a biological-inspired hexapod robot which contains a neural control system to perform basic locomotion. An interesting task for walking machines is manouvering through complex landscapes. Climbing over obstacles in various heights is a basic requirement to fulfill this task. Cockroaches are able to climb over obstacles higher than their own body height1 . Understanding the locomotion of climbing cockroaches is the key to develop a neural control system for climbing. The hexapod robot AMOS-WD06 is anatomically inspired by these insects. Thereby the neural control system of the robot consists of three modular networks working in parallel. Main locomotion is performed by a so-called Central Pattern Generator (CPG) which, with respect to wiring, provides a tripod walking gait. Two connected modules for Phase Switching (PSN) and Velocity Regulation (VRN) enable turning as well as omnidirectional motion. Different sensors control the reactive behavior of the robot. Simulation results analyzing the effects of these network modules are shown at the end followed by a brief outlook for the adaptive climbing behavior of the robot.


cated at the end of dendrites and membrane of the neuron. The activation of the neuron occurs when the signals surpass a certain threshold. After the neuron is activated, it emits a signal to other neurons. In order to simplify this complex network, it is sufficient to model artificial neuron networks (ANNs) in a highly abstracted way. These networks basically receive weighted input data (e.g. synapses). The activity of the neuron is determined by a mathematical function. A transfer function computes the output of the artificial neuron. ANNs combine artificial neurons in order to process information. Assuming that the activity of a neuron is given by the sum of all weighted signals of the previous time step: ai (t) =

wij σj (t − 1)

,i = j


whereas the output signal is a transfer function value of the activity of the neuron: σj (t) = f (aj (t)) = tanh (aj (t))
1 a (t) 1+e j


e.g. the hyperbolic tangent (first) and the standard sigmoid (second) transfer function. Furthermore, neurons can transfer their activity through a threshold function as shown in sections 4.1 and 4.5. Walking machines are mechanical vehicles (mechs or jpn. mecha) which perform bi- or multipedal locomotion. There are plenty of applications for walking machines. Many projects are intended for scientific, industrial or military purposes. Using a control system based on neural networks is advantageous to guide through rough terrain. Particular advantages are essentially due to its nonlinear dynamics. Thus, complex processes can be recreated without further knowledge of the physical relations or constraints. Further advantages are adaptability and fault tolerance related to processing imprecise, even deficient data.

Keywords: Cockroach locomotion, neural networks, hexapod walking, backbone joint control.



An autonomous system which can interact with its environment requires a complex control system. A large number of input data causes the system to react the ”right” way. One possible solution to these given problems assumes the approach of divide and conquer. Thereby complex problems are divided into simpler elements to be solved. In reversion, these solved problems can be decomposed into the solution of all. Networks can apply this approach whereas the nodes of the network can be seen as computional units. The information they receive gets processed and finally sent out. Connections between the single nodes determine the information flow. Each connection has a certain strength whose signals are preferred for conduction. One way to create an autonomous system is to observe natural neural networks and to adapt this knowledge. In this case, the nodes are neurons which receive signals through synapses lo1 [L1]. 2 [L1]


Locomotion of Cockroaches

Due to the fact that walking machines are adapted from biological knowledge, it is helpful to understand the climbing locomotion of certain hexapod animals e.g. insects. Watson et al.2 observed the locomotion of the deathhead cockroach (Blaberus discoidalis). Whilst climbing, the body requires to elevate its center of mass (CoM) to reach a higher altitude. This is done by a positive change of the body-substrate angle which is defined as angle between ground and torso axis. Comparing


Precisely. the PSN and the VRNs provide a certain steering capability.2007-11-12. The VRNs. All in all.climbing behavior over obstacles of different heights. After the tarsi touches ground. By and large. One leg pair is fixed at the front. Each leg has three joints for locomotion: The thoraco-coxa (TC) joint controls forward/backward motion of the leg. two ultrasonic sensors at the front of the trunk. The rear part has two leg pairs. it is observed that climbing does not require any remarkable dispositions from general walking control mechanisms. The neural control unit generates signals for locomotion of the robot. AMOS has 58 sensors: nineteen angle sensors for each joint. The physical engine is based on the Open Dynamics Engine (ODE).2008-02-04. A backbone joint connects the two parts of the torso of the robot. information from these simulations is helpful to be adapted to the actual robot system. This phase is named rising phase. see [I3]. The programming structure for robots is determined by a robot file. • a Phase Switching Network. This is done through leg extension of each leg pair. These networks are generally described in the following sections (see [L2] for a detailed view).g. supply the TC joints. Cockroaches climb over obstacles smaller than the height of their front leg swing trajectory with no change of walking gait. reacting to an obstacle the robot requires divisive sensory feedback (e.informatik. Whereas. for software development and project synchronization Eclipse (in addition with SVN) is used. its modular system structure is helpful for modifications. a simulation file and a controller file. For more information. The anatomy of AMOS is based on the insect body.2966424558/nldproject.ds. visual or antennae) as well as a neural control mechanisms for the 6 See: http://www. the coxa-thorachantal joint extends to push the front body from the ground. For higher obstacles the letter is anticipated before the front legs reach the obstacle. The front legs reach the obstacle’s edge through walking. there are three modules in use (Fig. LpzRobot4 is a selection of programs to simulate autonomous machines developed by a collaboration of the research groups for Neuroinformatics and Robotics and Self-organization in Adaptive Systems (University of Leipzig5 / MPI for Dynamics and Selforganization Göttingen6 ). robots need to receive several input data. To be able to fully climb the obstacle. Thereby. 3 [L4] 4 Online documentation: [I1] 5 See: http://robot. This anticipatory change of behavior moving the torso upwards is called rearing phase3 . 4 3 Bio-inspired Hexapod AMOS-WD06 Neural Control Unit AMOS-WD06 is a bio-inspired hexapod robot developed to study the neural control of locomotion of living creatures. 1): • a Central Pattern Generator. This leads to a positive change of the body-substrate angle.cpp. but an anticipatory behavior in the rearing six infrared sensors located at the front of each leg. six reflex ultrasonic sensors at each leg. Besides nineteen motors. the CPG contributes a periodic signal providing the leg rhythm of the robot. For details of the simulation of AMOS-WD06. Traversing the PSN the CPG signals reach the CTr and FTi joints.2710681714 2 .1 Robot Simulation Platform LpzRobot It is of advantage to create applications simulating robots in a certain environment without physically depending on the real machine. a difference in approaching this positive change occurs. In this case. • two Velocity Regulation Networks in parallel. which receive an output signal from one neuron of the PSN and else sensors. The neural control unit is located in the controller file amosIIcontrol. see [I2]. six foot contact sensors. the CPG sets the rhythmic movements of the legs. a translation of the CoM above the obstacle is required. 3. In addition.nld. the coxatrochanteral (CTr) joint has the role of extension and flexion of the second limb and the motion of the third limb (up and down) is driven by the femur-tibia (FTi) joint. The programming language of LpzRobot is C++. To respond to a given environmental situation. nineteen torque sensors for each joint and three 3daccelerometer sensors (x/y/z) at its body.

the outgoing motor signals need to be delayed.0 -1. In general. Input values for each direction7 are shown in table 1.0 1. these inputs affect the velocity of the right.5.4 Infrared Sensors The Phase Switching Network is a feed-forward network. such as an obstacle placed in front of the 3 .3 Velocity Regulation Networks The Central Pattern Generator is a recurrent two-neuron network. The PSN also reverses the phase of the periodic signals controlling the motors at the CTr and FTi joints.0 1. the input signals I2 .2 Phase Switching Network Table 1: Input parameters for the dierent walking directions and behaviors 4.0 1.0 0. The impact of the VRNs on the locomotion depends on the magnitude of the inputs I4 and I5 . This can be achieved by creating a buffer for previous signals. are counted to calculate the slope of the triangle. which.0 1.0 0.Figure 1: Neural control system of AMOS 4. 1).0 0.0 1.0 -1.and left-sided TC joints by scaling the amplitude of all signals. forward/backward locomotion is obtained as well as turning in each direction. modifies the phase between both output functions from − π to π . I4 and I5 influence the locomotion of the robot.0 1. Consequently.0 0.0 -1.0 0. Since the calculation is subsequent.18 + µ) where µ is a control variable for following experiments (see § 5.0 -1.0 0. Setting the rule σstep = 1 −1 if σ > σthr else The Velocity Regulation Networks control TC joints on each side.0 0.0 1. As previously mentioned.In accordance with this postprocessing.0 -1.0 0. I3 . 7 [L2].1). The weights of the connection between both neurons are selected ±k = ±(0.0 1.0 -1. Hence.0 1. 271 Infrared sensors are required for the anticipation of perception in certain situations.0 0.0 1.0 1. p.0 0. This result can be achieved by setting a threshold σthr for the CPG output.0 1. Recurrent weights wii are set to 1.0 1. See [L2] for more information on using the VRNs.0 0.0 0. recurrent means that the neurons can affect themselves.0 1. In this buffer system each following joint (starting with the rear right and ending with the front left) is delayed by τ = 16s each.0 -1. Postprocessing the signals of the CPG shapes the output function to a triangle form.0 0. each slope is predicted by the calculation done in the previous phase.0 0.0 -1. It consists of 2 2 12 neurons (see fig.0 1.0 1.0 1. leads to a step function. where the (step) function σstep is constant.0 0.1 Central Pattern Generator 4.0 1. Using both modules.0 4. The conversion of the step function σstep into a triangle function σtri is done by accumulation. according to a given input signal I3 .0 1.0 1. all time steps t. By this time. Eect Forward Backward Rightward Leftward Right turn Left turn Marching FDR FDL BDR BDL I2 I3 I4 I5 1. PSN and VRNs allow omnidirectional locomotion.

Experimental values are shown in table 2.1). ν2 : aBJC = ν1 (σFR + σFL ) + ν2 (σR1 + σL1 ) (4. This challenge can be accomplished by shifting the signals of CTr and FTi joints.4 and ν2 = 0. To elevate the CoM. a height calculation is helpful to determine the magnitude of the signal driving the backbone joint. It activates when the infrared sensors detect an obstacle of certain height. Hence. Preprocessed IR signals (IRR1 . each value for different walking gaits can be determined. the locomotion slows down.robot. IRFL ) are summed up with reference to each weight ν1 . The experiments show how the neural control unit reacts under certain conditions.5 Backbone Joint Control The walking gait depends on the frequency of the given CPG signal.2) 5. This is done in the simulation file main.1 Walking Gaits 4. The time of activation is marked further. The wiring generally provides a tripod walking gait with certain values of the CPG weights. 5. Hence tetrapod and wave gait forms are possible to deploy. the results of the activation by backbone joint control are described as well as the impact of given parameters on the locomotion of the robot. The joint activation induces a state which persists as long as the IR sensors of the middle leg pair receive no significant signal. It is shown that the joint signal frequency arises with increasing control value µ. The implementation of the rising phase has to be considered.h. the robot performs its regular walking locomotion. The range of the sensors is adjusted to the real values of AMOS. IRL1 . By variation of the variable µ (see § 4. Lowering these weights. learning of the estimation of a proper distance x the robot begins to climb with respect to different walking velocities and gaits. 4 . Furthermore.18 0. When the activity of the input neuron of the BJC trespasses the threshold value λ = 0.34 Walking Gait Slow wave gait Slow tetrapod gait Tetrapod gait Tripod gait Faster tripod gait Table 2: Walking gait values The control variable can be modified via console during simulation. a countermovement of the backbone joint needs to be anticipated. an obstacle is placed in front of the robot.cpp. Setting a threshold for activation ensures a distance x where the robot begins to climb. Up to a certain point. the control of the backbone is activated. whereas ν1 = 0. In addition. 2). the results of § 5.1) µ 0. the legs perform higher swing trajectories while walking.2. Besides that.02 0.14 0. The BJC is a proper solution to adapt the anticipatory rearing behavior of Blaberus discoidalis.05 0.3 give a brief outlook how to realize climbing behavior. Therefore. Fig. The results are pictured in figure 3. These sensors are located at the front of each leg and displaced by one third of the limb length downward from the center. Two additonal sensors are located at the front of the trunk. movements of the TC joints are captured as a function of time. IRFR . In order to reach higher grounds the use of the backbone joint is important. 4 displays the backbone joint motor signals as a function of time. The Backbone Joint Control (BJC) is an additional module to the neural controller (see Fig. Aforementioned tasks are deployed within the bachelor’s thesis.2 Activation of Backbone Joint This section points out the effects of the BJC. To compare each gait.5. This calculation may be used for adaptive mechanisms of the system such as decision behavior in ”obstacle situations”. These adjustments are done in the header file of the robot amosII. This is the activity of the input neuron of the BJC which transfers the activity via threshold: σBJC = 1 0 if aBJC > λ else (4. 6 Outlook Figure 2: Backbone Joint Control (Scheme) 5 Experiments and Results Henceforth. This behavior can be activated by significant signals from the IR sensors of the middle leg pair.

5 Output signal σ 0 -0.5 -1 30 32 34 Zeit t 36 38 40 Figure 4: Activation of the Backbone Joint 5 .1 CR1 0 -1 0 1 CL0 0 -1 0 1 CR0 0 -1 0 30 60 90 120 Zeit t 150 180 210 240 30 60 90 120 150 180 210 240 30 60 90 120 150 180 210 240 Figure 3: Leg movements with regard to variation of µ in time BJC activation Activation Backbone Joint Signal BJC Input Neuron Activity Threshold 1 0.

Bangkok.on.Backbone Joint Control (Scheme) Own work (Inkscape) 3 . L. et al. A.5 http://robot. M. 34 (2006): pp.Activation of the Backbone Joint Own work (Gnuplot) References [L1] Watson. S.: Sensor-Driven Neural Control for Omnidirectional Locomotion and Versatile Reactive Behaviors of Walking Machines Robot.. Online Documentation of lpzrobots 0.: Climbing Obstacles via Bio-Inspired CNN-CPG and Adaptive Attitude Control International Journal of Circuit Theory and Applications. T. December 1-5 ICONIP’09 (2009): Part II. E. Zill. Pasemann. J.informatik.2) LINK [L2] [L3] [L4] [I1] [I2] [I3] 6 .: LpzRobot Manual for Hexapod AMOS Simulation (Ver.html Open Dynamics Engine Official Site http://www. LNCS 5864. Patané. M. R. Pavone..5-doc/html/index. 1..ode. pp.. Pollack.. P. 1. F. 56(3) (2008): pp.. 47-55.. p. 109–125. Wörgötter. P. L.Appendix List of figures: 1 .. Manoonpong. Wörgötter.. Auton. F. Syst. P.. Frasca.: Control of obstacle climbing in the 265-288 Manoonpong.: Adaptive Sensor-Driven Neural Control for Learning in Walking Machines 16th International Conference on Neural Information Processing. N. Blaberus discoidalis: I. Ritzmann. 39-53 Manoonpong. F. J.Leg movements with regard to variation of µ in time Own work (Gnuplot) 4 . Fortuna. Locomotion J Comp Physiol (2002): Arena.html 2 ..Neural control system of AMOS http://www.