next up previous
Next: About this document ...

Hardware:
A modified commercial robot (``rug warrior'', $16~cm$ diameter) was used which is driven by two DC motors. Average speed was adjusted to $v=0.45~m/s$. Mechanical contact is detected by three microswitches $CS_l,CS_r,CS_b$ in triangular configuration. Visual information is aquired with two active infrared range finders $VS_l,VS_r$. The detection range was adjusted to $0.5~-~15.0~cm$. Interfacing between robot and computer is done tethered via a standard parallel I/O card and an industry A/D-D/A card. The C++ program ran on a Linux system in the real-time mode.

Sensor characteristics:
Sensor signals are band-pass filtered by the transfer function $h(t)$. The transfer function in the LAPLACE domain is defined as $H(s)={1\over (s-p)(s-p^*)}$ where $a:={\mathrm{Re}(p)}=\pi f/Q$, $b:={\mathrm{Im}(p)}={\sqrt{(2 \pi f)^2 - a^2}}$. The output functions of the band-pass filters are denoted as $u_i$ and normalised to one. The band-pass characteristics of all collision sensors $h_r(t)$ are identical and set to $Q=0.6$ and $f=1~Hz$. The signal from each vision sensor is fed in parallel into a filter bank of ten band-pass filters which frequencies are set to $f_i={{10}\over{i}}~Hz,~i=1,2,\ldots 10$; $Q$ is set to $1.0$, throughout.

Neuronal Circuitry:
The robot has two neurons; one which controls the speed $ds$, the other the steering angle $d\phi$. Normal operation is straight forward motion ( $ds=const,~d\phi=0$). Both neurons receive inputs from all sensors in a direct feed-forward connectivity.

Unconditioned retraction reaction:
The unconditioned retraction reaction uses only the collision sensor signals. These signals drive the output neurons in such a way that an avoidance movement with a motion vector pointing away from the site of stimulation is elicited.

Learning:
All band-pass filter outputs $u_i$ from the collision- and the vision sensors converge onto both neurons where they are summed according to their synaptic weights $\rho_i$. Only the synaptic weights from the vision-sensors are allowed to change starting from zero. The change of the weights is achieved by the learning rule defined as: $\Delta \rho_i = \mu u_i
f(v) := \mu u_i v^\prime$ where $v$ is either $ds$ or $d\phi$. The constant $\mu$ is set to $0.00002$.

Neuronal Output:
The output of the neurons is defined as: $ds=v-\rho_0^{ds} [h_r(t)\ast (CS_l+CS_r-CS_b)]+l_{ds}$ and $d\phi=\rho_0^{d\phi} [h_r(t)\ast (CS_l-CS_r)]+l_{d\phi}$. The asterisk denotes a convolution operation. The variables $l_{ds}$ and $l_{d_\phi}$ represent the total sum of all learned contributions that converge onto the $ds$- and $d\phi$-neuron respectively. The synaptic weights in unconditioned reaction are kept constant at $\rho_0^{ds}=0.15$ and $\rho_0^{d\phi}=-0.5$.




next up previous
Next: About this document ...
Bernd Porr 2001-07-16