Machina Speculatrix
schip notes -- v0.2

M. speculatrix are robotic tortoises or turtles -- tortoise seems to be the original but now turtle has taken precedence -- originally designed and built in l948-49 by W. Grey Walter, a British neurologist at the Burden Neurological Institute in Bristol. They comprised two motors, two vacuum tubes, four relays, a bumper switch, and a light sensor. They could autonomously wander, follow a light, and find their re-charging station. He published two articles about them, (Walter, W. Grey, "An Imitation of Life," Scientific American, May 1950, p42-45), and included them in his book (Walter, William Grey, "The Living Brain," Duckworth, London, 1953). During the 1950's they were a hit at various technology exhibitions in Britain and the USA. I suspect that one was lost in shipping to the 1970 Software show at the Jewish Museum in NYC, but I can't prove it -- I found no record of Walter being included in any of the draft lists of artists for the show in the Museum archive (one 40 year old box of leftover papers).

Here is the physical layout of the first of the M. speculatrix race, from the cyberneticzoo.com page:

elsie layout

This is the cleaned up and annotated schematic from the beam-wiki.org page:

original schematic


time lapse
Above is a time lapse photo of Walter with Elmer and Elsie (Electro-mechanical robot, Light Sensitive with Internal and External stability) from a 1950 Life magazine article (courtesy of cyberneticzoo.com).

And an early news film (from YouTube http://www.youtube.com/embed/lLULRlmXkKo) of them wandering around ...sorry about the soft-porn cartoon music...:



Behavior

The M. speculatrix behaviors resulted from a serendipitous mechanical system where the motors did not reverse -- the steering head had to rotate 360º in order to come back to moving straight forward but the robot would move backwards during 180º of that rotation -- and careful selection of time constants in the simple "neural" circuit formed by the tubes and relay coils. In my effort to build a modernized model of these turtles I have extracted the behavorial descriptions from the beam-wiki.org page and mapped them to the (¿simplified?) state description below. I am not attempting to fully emulate Elsie. My GWT model can only rotate the steering head by 180º but it has the same effective range of motion because it's motors can be reversed. In addition to the basic light sensor I added [a PIR sensor so it can find people as well as light *** this does not work as the sensor I used triggers whenever it moves regardless of of the {ab,pre}sense of a person ***] dual right and left hand bumpers, and the ubiquitous Arduino in place of the simple complexly-tuned tube and capacitor brain. With this setup my GWT can be a bit more sophisticated when tracking signals. The originals also had a headlamp which could be the light source tracked by another of their kind. I have replaced the PIR sensor with a pilot LED which may be used for future tracking, but I'm only making one robot for now.

Walter also makes a point of describing the photocell as being wired so as to make it sensitive to changes in intensity rather than just fixed levels, and thus able to adapt to its environment in some small way.

operating states
original circuit description simplified description schip's GWT operation
SEARCH
This is the normal operating condition of the circuit with Tube 1 biased ON and Tube 2 biased OFF. As a result Relay K1 contacts connect half current to the Drive motor and Relay K2 contacts connects full current to the Turn motor.

The Steering motor continuously turns the vector of the front drive wheel and the photo cell. The result is a spiral motion of the turtle with the aperture of the photo cell scanning the horizon like a rotating radar dish.

When the turtle returns from MOVE to SEARCH mode the Photocell turns away from the light source, the Tube 1 control grid voltage rises and turns Tube 1 further ON. The negative pulse on the Tube 1 anode coupled through C2 further turns Tube 2 OFF.

The Photocell continues to turn past the light source. As it rotates fully, the same light source, now closer, comes into view again and the turtle returns to MOVE mode.
When there is no signal from the Photocell, the Steering motor rotates at full speed while the Drive motor rotates at half speed resulting in slow small spiral movements of the turtle.

If the Photocell detects a signal the robot goes into the MOVE mode.

When the turtle returns from MOVE to SEARCH mode, the Photocell will have rotated away from the light. The turtle continues rotating until light is again detected. 		With no input signal the GWT runs it's Drive motor at a slow speed and sweeps it's Steering motor 180º back and forth over a period which allows it to make a full circle. This scans the horizon as it were.

If a signal is detected the GWT goes to the MOVE mode.

If a bumper is detected the GWT goes to the TOUCH mode.
MOVE
When the Photocell is exposed to a moderate light level it partially discharges C1 which lowers the Tube 1 grid voltage. This dip in grid voltage is amplified by Tube1 and appears as a positive output signal on Tube1 anode. Note also that the input signal is small and Tube 1 anode current remains above the Relay K1 drop out current.

This positive output signal is capacitively coupled through C2 to the grid of Tube 2. If this signal exceeds the turn-on threshold of Tube2, it is further amplified and energizes Relay K2 coil.

Relay K2 contacts change state and the Steering motor stops while the Drive motor is connected to full current. With the photocell rotation stopped pointing to the light, the turtle now moves in a straight line in the direction of the light source.

This MOVE motion continues for the duration of the R4/C2 time constant until the voltage at the Tube 2 grid drops the anode current below the Relay K2 holding current. Relay K2 contacts reconnect the Steering motor to full current and the Drive motor to half current.
When the Photocell detects a signal, the turtle goes from SEARCH to MOVE mode.

The Steering motor is stopped and the Drive motor rotates at full speed. This causes a motion straight towards the light.

When [(I think...) the Photocell no longer detects a signal AND] a short timeout period has elapsed, the turtle returns to SEARCH mode. 		With a signal detected the GWT runs it's Drive motor at full speed and tracks the signal by alternating the direction of the Steering motor to maximize the signal. It moves towards the signal.

If the signal exceeds a pre-determined level the GWT goes to the DAZZLE mode.

If a bumper is detected the GWT goes to the TOUCH mode.
DAZZLE
This occurs when the light level on the photocell is high enough to deeply discharge C1. Then Tube 2 anode current drops below the Relay K1 holding current and Relay K1 contacts open to connect half current to the Steering motor. At the same time the positive transition on the Tube 1 anode turns on Tube 2 which causes Relay K2 contacts to connect the full current to the drive motor. This is the first part of the DAZZLE mode in which the Drive motor rotates full speed and the Steering motor rotates half speed which causes the turtle to veer away from the light. The turtle moves rapidly away from the light in a slow curve.

When the Photocell detects a strong signal the turtle enters DAZZLE mode. The Steering motor rotates at half speed and Drive motor at full speed.

The turtle moves rapidly away from the light in a slow curve.

As the Photocell is slowly turned away from the light it's signal drops and the turtle enters BLIND-SEARCH mode. 		With a strong signal detected the GWT runs it's Drive motor at full speed and turns the Steering motor fully to the opposite direction of its last travel.

This lasts for a period allowing a rotation of between 45-90º. At the end this period the GWT goes to BLIND-SEARCH mode.

If a bumper is detected the GWT goes to the TOUCH mode.
BLIND-SEARCH
This is the second phase of the DAZZLE mode, which occurs when the photo cell turns away from the light and C1 charges up again. When the Tube1 grid voltage rises it turns Tube 1 ON and Relay K1 changes state. Importantly, the large negative transition on Tube 1 anode, coupled through C2, turns Tube 2 OFF which drops out Relay K2. This large negative signal drives Tube 2 deep into cutoff and ensures that during the following few seconds Tube2 will not respond to small positive pulses on the Tube1 anode resulting from periodic illumination of the spinning photocell.

This is the same relay logic and motor operation as the SEARCH mode (Steering motor rotates at full speed and the Drive motor at half speed) but photocell signals are ignored until Tube 2 returns from cutoff.
When the Photocell signal drops from a high value the turtle goes from DAZZLE to BLIND-SEARCH mode. This persists for several seconds, during which time the Photocell signal is ignored.

As in SEARCH mode with the Photocell signal blocked, the Steering motor rotates at full speed and the Drive motor at half speed, resulting in slow small spiral movements of the turtle.

When the Photocell-ignore timeout elapses the turtle returns to SEARCH mode.

The net result of DAZZLE and BLIND-SEARCH is a fast large spiral away from the light source followed by a period of random small spirals until the normal phototropic SEARCH/MOVE behavior can resume.

Tracking input signals are inhibited for a period of a few seconds while the GWT runs it's Drive motor at a slow speed and sweeps it's Steering motor 180º back and forth.

At the end of the inhibit period the GWT returns to SEARCH mode.

If a bumper is detected the GWT goes to the TOUCH mode.
TOUCH
The TOUCH mode occurs when the collision contact closes, which connects Tube2 output to Tube1 input through C3, R7 and R8 forming a cross coupled oscillator. The two inverters are connected in a loop through differentiating R/C networks and this positive feedback causing the circuit to oscillate between SEARCH and DAZZLE modes. The duty cycle of oscillation is 1/3 but is affected by the light level. (In BEAM robotics this network connection would be called a two Nv neuron grounded bicore).
When the bumper contacts are closed the turtle alternates between SEARCH and DAZZLE modes with a duty cycle of about 1/3. The period of this depends on the Photocell signal, such that oscillations in the dark take longer than in light.

This alternates Steering and Drive motors rotating at full and half speed resulting in slow small spiral movements of the turtle interspersed with fast open curves. And (hopefully) it moves away from the obstruction. 		When a bumper is detected the GWT runs it's Drive motor in reverse at a slow speed and turns it's Steering motor towards the bumper switch that was activated. If both bumpers are activated the GWT turns to the opposite direction of its last travel. This backs up and turns away from the offending bumper.

TOUCH mode lasts for a period to allow a rotation of about 90º, at which time the GWT returns to SEARCH mode.
LOW-BATTERY
The LOW BATTERY MODE can occur when circuit bias conditions are chosen so that a low motor battery voltage will result in the MOVE mode without the DAZZLE termination and hopefully results in the turtle entering its brightly lit hutch for a battery charge.

The battery charger interface cleverly uses two additional relays to disconnect the motor battery from the rest of the circuit to protect the battery from over discharge, to turn of the motors and the rest of the circuit during charging and to reconnect the battery to the circuit when fully charged.

As long as the charging current is high enough, the K4 contact remains open and charging continues.

When the battery is fully charged, the charging current tapers off and K4 drops out. This reconnects the positive 6V battery terminal to the main circuit which powers up M. speculatrix to drive forward and immediately enters the Touch mode to extracts itself from the charging dock and resume normal mode operation.
When the battery is low DAZZLE mode is disabled and the turtle continues to MOVE towards the light. If the light is at a charging station this will (maybe) cause it to connect and recharge.

While charging the drive system is disabled.

When the battery is fully charged the normal SEARCH/MOVE behavior is re-enabled and the turtle moves away from the charging station's bright light. 		When the battery is low the GWT will turn it's Steering motor fully to 180º and run it's Drive motor at slow speed alternating directions as a signal that it needs juice.

Input signals and bumper switches are ignored.
CAPTURED-DAZZLE
If there is a possibility that the photocell is continuously exposed to bright light no matter what direction, the turtle can hang up with the steering motor ON and the drive motor OFF.
I'm not sure how this happens???

Or what it means... 		If there is no signal change detected, either consistently too low or too high, the GWT remains in SEARCH mode.

Further Study

Walter made a second robotic attempt using behaviorist learned behavior in what he called CORA -- COnditioned Reflex Analog, or M. docilis. He describes this in The Living Brain chapter 7, The Seven Steps from Chance to Meaning and details the circuit in Appendix C. The brain was expanded to seven tubes with an added sound sensor. It was, with careful tuning, capable of learning a conditioned response by associating sound with a series of physical bumps. There is some indication that this circuit was retrofitted to one of the Turtles. However the first paragraph of the appendix says: "It was worked out to fit on to M. speculatrix, but it is difficult to adjust on a moving model and should be set up on the bench first." There is a good cyberneticzoo page which tries to decipher exactly what and when this all happened. That page also attempts to explain the turtle/tortoise nomenclature by applying turtle to M. speculatrix and tortoise to M. docilis.

Links