Autopod 0.33, an untethered pneumatic walking metamere robot

by Jan Henri Cocatre-Zilgien
(C) 1996

Back to Jan Cocatre-Zilgien's homepage

1. Introduction

Walking robots present challenges that go beyond the mere coordination of walking, which is widely perceived as the main problem to be solved for such machines. A walking machine needs to be a harmoniously integrated system of general design, leg design, actuators, actuator controls, energy and power supplies, external and internal sensors, navigational equipment, communications links, and controlling system. The building of a successful machine is more likely to be hampered by the mismatch between all of these components than by any one element in particular being ill-suited to its task.

The Autopod project consisted in the design, the component testing, the construction, and the evaluation of a proof-of- concept walking robot, within one year (1996). Its purpose was to gain more insight on the problems related to the interface and integration of multidisciplinary systems. Unlike the building of a new car, for which hundreds of models exist, there are no guidelines for the building of a walking machine.

Objectives were not defined in terms of shape, materials, or operating principles, but in terms of function. The function chosen for Autopod was a useful and real one, namely Search and Rescue in an urban environment. A typical mission would involve carrying a TV or an infrared camera to find survivors trapped in collapsed and unstable buildings after earthquakes or explosions. For example, a large and heavy slab of concrete hanging by a single reinforcing bar prevented rescue personnel from reaching an area of the federal building recently destroyed by terrorist bombing in Oklahoma City (and the slab actually did spontaneously crash down a few days later). Another typical mission would be exploring an area of toxic or explosive gas, such as in a mine or an accidental spill of the chemical industry, which would prevent the use of search dogs.

The mission of Autopod therefore imposes the followings constraints on the robot: It has to be relatively small and easy to bring to the location of a disaster area. It has to be of low acquisition and operational cost, so that although not designed to be expendable, it can be sent without hesitation in dangerous areas from which it may not come back. It has to be robust and untethered, given the terrain it has to tread. It has to be simple, fast to set-up on site, and easy to operate by mostly untrained personnel, often in difficult and urgent conditions.

The development of Autopod led to an hybrid intermediary machine having two legs and two wheels, dubbed "Autopod 0.33" (front cover photo), essentially as a third of a six-legged machine end product. This version maintained the need of system integration, and allowed exploring the biological concept of "metamere" for control. Autopod 0.33 uses inflatable instead of air cylinder actuators, controls its actuators by novel kink valves instead of solenoid valves, and is powered by a high- pressure tank of compressed air instead of an on-board compressor. Given the partial limitation that this robot is a hybrid leg-wheel arrangement, the objectives were reached, except performance-wise on slopes, for reasons which will be discussed later. [return to top of document]

2. Robot structure

2.1. Body structure

The body of the robot is made entirely of L section aluminum (12" x 12" x 1/16"), bolted together in a very rigid interlocking frame that does not need bracing (Fig.1). The frame is hollow and of easy access and supports directly, from the top down, an electronics board, a valves board, low-pressure tanks, and a high-pressure tank. The top (electronics board) is protected by a hemicylindrical sheet of clear polycarbonate.

Figure 1. Three-quarter front view photograph of Autopod 0.33 showing
the body frame, the longitudinal offset of the wheels relative to the
body, and the legs of the metamere.

2.2. Leg structure

Each leg of the robot is made of aluminum tube (3/4" OD, 1/16" wall) and square-section aluminum tube (1" side, 1/16" wall). It is essentially a pantograph leg, so that the tibia stays parallel to the coxa (usually vertical) when extensors and flexors of the tibia are equally activated (Fig.2). This leg with all apparent - and vulnerable - parts was designed for ease of repositioning pivoting axes and pulling points; the leg of a more robust robot would be designed after the Morrill Leg (Dept. of Entomology, University of Illinois at Urbana Champaign), that is, with a riveted steel loadbearing exoskeleton.

Figure 2. Front view photograph of Autopod 0.33 showing the right leg of
the metamere bearing weight on the ground and the left leg lifted up.

2.3. Hybrid leg & wheel arrangement

The walking metamere is essentially a pair of legs from a 6- legged machine, with the four missing legs replaced by a pair of low pressure (15 PSI) 14" diameter unpowered wheels. The wheel axle is long enough that the robot is always tripod-stable with any one leg and two wheels on the ground. Location of the center of mass relative to the wheels is such that the legs bear approximately the same load as they would in a 6-legged version (see front cover photo and section 6.3 below). Wheel arrangement also allows load variation by adding weights above the location of the legs to simulate payloads, or by adding counterweights on the other side of the axles to partially simulate buoyancy or low-gravity. However, the wheels introduce problems of their own, especially on sloped terrain.

The metamere robot control is simpler, as the robot has only 3 degrees of freedom: pitch up & down (metamere elevate-crouch down), yaw left & right (metamere crab left & right), and movement forward & backwards. Except for some stability issues, integration of all the subsystems and results can reasonably be extrapolated to a 6 degrees of freedom 6-legged version. [return to top of document]

3. Motor actuation

For simplicity, weight, and compliance reasons, compressed air was selected over electricity as propulsive power. Also, compressed air is safer in explosive environments. Working pressure was 100 PSI, but without payload, Autopod 0.33 actually walked with 70 PSI, and the whole system has been tested at 160 PSI for "emergency power". Because this was a personal project, hydraulics were not considered for logistics reasons.

3.1. McKibben Actuators

Leg design is intertwined with that of the actuators. One of the problems that emerged during component testing is that mid-size cylinders, even the aluminum-ended ones, are somewhat heavy for the work they can produce under 100 PSI. A computer program was written to optimize actuator and leg geometry, as a function of actuator weights (which are not listed in catalogs and have to be extrapolated from samples & reconstructions). As an alternative, custom cylinders made out of rigid PVC tubing were built, but the bore of plastic tubing turned out to be irregular and mottled enough that no O-ring could slide in them smoothly. In the end, custom-made McKibben actuators were built and used successfully for all actuators of Autopod 0.33. It is apparently the first time such actuators have been used for a walking machine, although it is possible this was done in Japan, where McKibben actuators are manufactured and where there are many types of walking robots. As witnesses of the robot have been intrigued by the life-like appearance and the motion of these inflatable actuators, they will be described now in some detail.

3.1.1. History of the McKibben Actuator

In the 1950s, American physician Joseph L. McKibben invented a pneumatic muscle actuator of a deceptively simple design to power artificial arms in amputees. His actuator consists of a rubberlike inner tube inside a hollow braided sleeve; increasing the pressure of a fluid in the inner tube increases the diameter of the actuator, and shortens it by pantograph action on the braided sleeve. The invention seems never to have caught on, presumably because the amputees had to carry a pressurized tank on their backs. In robotics, however, these artificial muscles are much more appealing, and the Japanese corporation Bridgestone has patented its "Rubbertuator" based on an extension of the same concept. For convenience, Autopod actuators were custom built, after what developed into significant experimentation.

3.1.2. Materials and construction

The construction of the McKibben actuator described here is built with components that can be found in nearly any hardware store. As a result its cost is low, several times less expensive than an equivalent cylinder counterpart. The parts needed for a typical $4 one-foot long actuator are listed below:

Figure 3. Photos of two different custom-made McKibben actuators.
(A) The top actuator is of the kind used in Autopod 0.33. (B) The bottom
actuator was an unsuccessful attempt at reducing the number of braids to
increase pulling force, and inner tubing burst out between braids.

The assembly proceeds in two steps. The first step is the preparation of end plugs for the inner tube. One of the branches of the nylon L is sawed off, and the cut sections are plugged with machine screws secured with epoxy glue. A short length of brass tubing or a specific fitting is epoxied for the air supply hose in the remaining port of the sawed L. Epoxy is allowed to cure 24 hours. Supply hoses of appropriate pressure rating (generally polyethylene) are to be used. The second step is the assembly proper and takes but a few minutes. With a pair of smooth needle-nosed pliers, a lozenge-shaped space is stretched open between braids, the latex inner tube is inserted inside the rope, and then, by a strange operation akin to feeding a worm to a snake, it is pushed manually in until only the free branch of the L fitting protrudes from the hollow rope. Care must be taken so that the latex is not damaged by the tool. Finally, the two extremities are folded around a 4" diameter rod or other pivot and the clamps are screwed exactly over the plugs, which can be located by touch. It is better to position the worm screw of the clamp on the folded side of the loop so as to pad as much polypropylene as possible between the worm screw and the underlying latex.

3.1.3. Performance and characteristics

A test actuator was built as described above, with an active length of 30.48 cm (1') and a pivot to pivot length of 37 cm, to obtain some quantitative data and assess its performance. SI units are used below for comparison with the scientific literature. An unpressurized actuator was hung vertically and loaded incrementally by the addition of weights up to 728 N (163 lb) then unloaded incrementally down to zero load, at which the length was remeasured at 37.7 cm, defined as the zero-length. The same cycle was repeated with different pressures up to 689 kPa (100 PSI, 68.9 N/cm2) in increments of 138 kPa (20 PSI). The actuator is pictured in Figure 3A, and force-length results for different pressures are plotted in Figure 4. These results influenced the overall design of Autopod's legs.

Readily apparent from the graphs is that the force produced depends on actuator length (which is also the case of biological muscles). For a given pressure, for example of 689 kPa (100 PSI), the pulling force decreases with the shortening of the actuator, from one extreme of 550 N (123 lb) for no shortening at all (0%), to the other extreme of 0 N for a shortening of 8.5 cm (22%). By comparison, the force produced by a piston is independent of its position in a cylinder actuator. If emulating air cylinders was the objective, drawing rectangles within the boundaries of Figure 4 would allow to identify equivalent ones; for example, a 1.66 cm (21/32") bore 5 cm (2") stroke, as drawn in Figure 4.

The actuator diameter increases from 1.75 cm (0.69") deflated to about 2.6 cm (1.02") inflated. Dividing the zero- length force (550 N) by the inflated cross-section area (5.3 cm2) yields a maximum stress of 1030 kPa (103 N/cm2, 149 PSI). Note that it is stronger than a piston of similar diameter (689 kPa), and more than three times stronger than a biological muscle (350 kPa). At an air pressure of 500 kPa, maximum stress is 755 kPa, only 65% of the average of 1160, 880 and 1440 kPa gathered from the literature at that air pressure, but a good result considering that all the components are off-the-shelf and extremely robust.

Figure 4. Force-length relationship of the actuator at 5 different
pressures (see text). Hysteresis trajectory is clockwise. The rectangle
defines the performance envelope of a 21/32" bore 2" stroke air cylinder
used for comparison.

The force-length diagram also reveals that at a given length, the force produced is not the same depending on the direction of the change of actuator length at that time. This hysteresis effect, due to friction, complicates the task of the actuator controller because it is difficult to predict. This form of viscosity could be called a "spatial delay" in the operation of the actuator.

The weight of the actuator is 64 g (2.25 oz), that is four times less than an equivalent cylinder, but as two McKibben actuators are needed to replace one double acting cylinder, weight advantage is two times less, still a very significant advantage for a pneumatic walking robot.

3.1.4. Design considerations

One disadvantage of McKibben actuators (shared with Nitinol wire and some biological muscles) is their short stroke for their length, about 22% without load but more like 10% with a useful load, compared with the 40% that can be reached by some cylinder actuators. As a result, actuators may have to be positioned away from their point of action, just as finger extensor and flexor muscles are located in the forearm. Since they only pull, as opposed to double action cylinders, cables and pulleys (or teflon guides) can reroute the direction of the force produced.

Another limitation is that only one size of rope and inner tube is given here and the resulting actuator is unlikely to be adapted to a prespecified task, as opposed to cylinder actuators, for which a wide choice of diameters and strokes exists. However, the stroke can be extrapolated from the data in Figure 4; a muscle twice as long will provide twice as much stroke (actually slightly more, as the length of the two end loops is constant). Also, the force can simply be multiplied by linking several actuators in parallel, which is easy for reasons explained below.

McKibben actuators are robust and can be somewhat bent longitudinally and twisted axially, so that the placement of its end attachments is very flexible (unlike cylinders, which have to obey strict trigonometry). They can even "jump" joints if too long to fit in a limb segment. Another aspect of their robustness is that when stretched (Fig.4) they tend to become like the hefty rope they were in the first place, here with a tensile strength of 17 kN (3800 lb). As a result, they naturally act as "stops" preventing damaging joint hyperextension or hyperflexion. In a typical antagonistic extensor-flexor arrangement (Fig.2), the extensor should be able to shorten completely under 100 PSI pressure without fighting a stretched flexor under 0 PSI on the other side, and vice-versa.

Except for their low maximum stress relative to their published or advertised counterparts, these actuators did not have any operational problems (and never burst as in Fig.3B). There is obviously room for improvement of the performance by selecting a better braided sleeve than polypropylene rope and a slightly thinner walled inner tube.

3.2. Valves

3.2.1. Hinged Kink Valves

It was no surprise that valves to control the actuators of Autopod 0.33 were a problem, as experienced with Protobot in the Hexapod Group at the University of Illinois at Urbana-Champaign. The "leaky pulse frequency modulated system" developed for Protobot turned out to consume too much compressed air for an untethered robot, at least in a simple form, so it was abandoned. Commercial valves were too heavy, too power-hungry or too expensive. Quite a number of lightweight spool valves with a wide variety of plastic and metal tubing were experimented with, but either they exhibited too much stiction or they just leaked. They needed close tolerances that could not be attained simply, and this in turn impacted on their robustness, so spool valves were abandoned too. This need for a valve led to the "hinged kink valve" concept. At first its arrangement seemed too simple to be effective, but it worked so well first in extensive trials and then in the robot itself that it was a very good surprise indeed. Trials for Autopod valves were conducted to the thousands actuations, but similar kink valves in the literature can be opened and closed up to one million times without failure. This valve fulfilled all the requirements for the walking robot.

3.2.2. Valve patent application

The principle of kink valves goes back to at least 1892, but hinged kink valves seemed novel, so given its great advantages, a patent has been applied for (Oct. 23, 1996). The patent application summary and a drawing (Fig.5) are reproduced here: "A two-way flexible valve is provided to control or interrupt the flow of a medium. The valve comprises a semi-rigid length of tubing that has been kinked by bending it until the material has yielded in an irreversible manner. Ulterior flexions of the tubing then tend to take place about the preferred axis set by the kink. An acute angle will close the valve by crimping the lumen shut, whereas an obtuse angle will open it, in some proportion of the angle. One end of the kinked valve tubing is held by a fitting on a support, and the other end is held by a fitting fastened to a lever hinged on pivot pins. The distance between holding fittings and the valve kink itself is several times the diameter of the valve tubing, and the axis of the kink is collinear with that of the hinge, both factors reducing stress in the valve material and increasing its lifespan. An elastic member holds the valve closed, and a flexible actuating line acting against the elastic member opens the valve in some proportion of its pull. The valve is simple, inexpensive, robust, of low maintenance, and because of its semi-rigid material, can withstand pressures higher than most flexible valves. A three-way valve embodiment is detailed for the control of a robotic actuator."

Figure 5. Patent application drawing of the normally closed three-way
hinged kink valve embodiment. Some relevant parts are supply port 25a,
actuator port 43, exhaust port 25b (with optional silencer 48), servo arm
47a&b, flexible lines 29, rubber bands 27. Drawing shows valve from
supply to actuator open, and valve from actuator to exhaust closed.

3.2.3. Valve implementation

In the implementation for Autopod 0.33, the support is a cut and bent thin aluminum sheet, the lever is made of bent piano wire, and the elastic members are plain rubber bands. Valve actuation is created by the arms of a standard model aircraft servo (Fig.6). All valves and servos are secured or embedded in 3/4" thick expanded polystyrene, with one end for electrical connections, and the other for the actuator connections (Fig.7). A main manifold runs under the board to feed air to each three- way valve. The servos are controlled by Pulse Width Modulation (PWM). Air flows practically full bore from supply to actuator for a pulse width of 2 ms, and flows practically full bore from actuator to exhaust for a width of 1 ms. Between these two extreme values flow is throttled accordingly, in about 100 steps in each direction, with a complete closure of both two-way valves for a 1.5 ms pulse width (closed center mode). The relationship between kink angle and air flow was not investigated, but it is very likely non linear.

Figure 6. Implementation of three-way hinged kink valves on a styrofoam
board for Autopod 0.33.

Figure 7. Top front photo of the robot. Single extensors and flexors of
the left and right coxa are not powerful enough to climb slopes. The 12
compressed air lines from the servovalve board are clearly visible.

3.3. Compressed air power supply

Originally a small on-board gas-powered compressor was envisioned, because the excellent energy to weight ratio of gasoline would provide long endurance. A wide range of lightweight gasoline engines exist from 3 to 60 HP, for model or experimental aircraft, and all develop about 1 HP/lb. However, suitable compressors to hook up to them, except for some rare exceptions (see section 6.4), were not found. Compressors are usually designed without any weight consideration (using cast iron, for example) for fixed industrial plants and struggle at about 0.1 HP/lb. The requested autonomy depends on the mission of the robot. The mission of Autopod 0.33 was not to beat a record of endurance, so the idea of an on-board gasoline engine to power it was dropped. Also, a gasoline engine would not be suited to operate in air of low oxygen content or an explosive atmosphere.

Figure 8. Rear view of Autopod 0.33 showing, from top to bottom,
hemicylindric polycarbonate shield, electronics board with switches and
two serial ports, servovalve board, low pressure tanks level, high
pressure tank level with pressure regulator, and attachments for removable
axle & wheels unit. Quick-connect fitting for optional hose tether is
just outside of the picture.

As a consequence, actuators of Autopod 0.33 are powered either by a quick-connect 3/8" hose tether linked to an electrical 3.5 HP compressor for the prototype, or by 3000 PSI air tanks for the operational version. A 20 foot tether is used for indoor testing, and a 100 foot long tether is used for outdoors testing. The 3000 PSI tanks are high pressure cylinders made of fiberglass-wrapped aluminum ( SCI , Pomona, CA). They have a Department of Transportation exemption and need to be pressure proofed every 3 years. They are expensive and it takes months to obtain them, but the energy they contain is comparable to that of lead-acid batteries of the same weight. They can be recharged at Fire Departments and at commercial dealers of firefighting equipment ($15/cylinder). In an operational 6-legged Search & Rescue robot, these tanks can either be exchanged for a filled one in a few minutes, like huge batteries, or recharged, also in a few minutes, from the high-pressure supply available in nearly all Fire Trucks.

A standard and heavy (3 lb) single stage industrial regulator brings down the pressure from 3000 PSI to a manually adjustable range of 0 - 180 PSI (Fig.8). 3/8" diameter brass tubing serves as the main manifold on the whole length of the body. To minimize main manifold pressure variations, four one- liter Windkessel tanks are connected to it in parallel. These tanks are made of fiberglass-wrapped polyethylene bottles proof- tested to 160 PSI, and are of insignificant weight and cost. There was concern originally that the regulator would ice up and clog because of the high flow and the dew point of the typical humidity of Illinois, but this never occurred. During intensive use outdoors, both regulator and high-pressure tank get cold and covered with dew though, a good reason not to put any electronics below them. [return to top of document]

4. Sensors

The number of sensors per leg is the minimal required, four, and this of course pales in comparison to the thousands on a cockroach leg, but it is enough for quite complex types of robot behavior.

4.1. Leg sensors

The three joint angles are measured by 50 kn potentiometers and encoded by 8-bit Analog to Digital to Serial chips (ADC0831). Eight bit resolution (256 steps) is sufficient for a walking machine, as opposed to pick-&-place robotic arms.

Leg contact information is provided by heavy duty momentary push-buttons secured in a rubber foot at the end of the tibiae. They can be triggered up to inclinations of 45 of the tibia relative to flat ground. Their activation force is 4 lb and their deactivation force 3 lb, which creates an hysteresis that prevents the sensor from hesitating in describing the state of the leg. Initially a compression spring with a linear potentiometer encoding some analog value of the force was tried, but abandoned as not rugged enough. The sideloads at the feet are surprisingly destructive for foot sensors. A foot instrumented with a few strain gauges may be a better choice, if need arises.

4.2. Other sensors

Pressure in the main manifold is monitored by a Fujikura bridge-type piezoresistive pressure sensor ( Servoflo , Lexington, MA), amplified by a custom built differential instrumentation amplifier, and encoded again by an Analog to Digital to Serial chip. The amplifier works well enough on 9.6 V instead of the required 12 V, which is not available on-board Autopod 0.33. All the sensors are successively polled by the sensory ganglion (section 5.1.2 below) and transmit their eight bits serially to the latter. More advanced multiplex chips are available for that task. Planned but not purchased for Autopod were a pair of two- channel 900 MHz video camera transmitters ($300 each), for demonstration of remote stereoscopic viewing - and Searching - by human observers. [return to top of document]

5. Control

5.1. Hardware

The hardwiring of the metamere generally follows the organization and terminology of biological systems. The nervous system consists of an upper brain, a sensory ganglion, a brain, and motor ganglia, organized in a one-way linear fashion, as shown in Figure 9. Note that the brain of this robot cannot modify sensor activity, only interpret it. Since Autopod 0.33 is comprised of only one metamere, a functional pair of legs, the nervous system of that metamere is fused for simplicity with that of the robot as a whole. It would be different in a three- metamere robot. All the electronics are soldered on modular perforated boards or printed circuits that can be easily reshuffled and debugged. They can be organized in a much more compact way if needed. All boards are held by screws on a same support made of 3/4" expanded polystyrene. With standard grounding methods, static electricity was not a problem.

5.1.1. Upper brain

Autopod 0.33 does not possess any upper brain per se, that is, it is unable to exercise its own free will. It takes all of its higher orders directly from a human being via the joysticks of a 7-channel 75 MHz radio control ( Futaba , Irvine, CA). Because of the wheels, Autopod 0.33 has only three degrees of freedom, controlled independently from one another by three channels:
1) walk forward or backwards,    (metamere walk),
2) yaw left or right             (metamere crab),
3) pitch up and down             (metamere elev.-depress.).
A 6-legged robot would possess the full six degrees of freedom of any physical object in 3-D space, that is, added to the above:
4) walk crabwise left or right          (yaw in 0.33),
5) elevation or depression            (pitch in 0.33),
6) roll left or right            (impossible in 0.33).

Not having an upper brain does not make Autopod 0.33 fully autonomous in the sense that if released in an unknown environment it could make strategic decisions as to which direction to proceed or which attitude to adopt. As a matter of fact, a fail-safe system automatically shuts down the robot in the event of the loss of the radio signal. However, given those motion requests, the robot would follow them without falling. The six channels (three in Autopod 0.33) of higher orders along this format of three translations and three Euler angles are convenient and natural to transfer motion requests, irrespective of the platform used. Any man-out-of-the-loop upper brain can simply be grafted onto the robot if its output has this same six channel format.

5.1.2. Sensory ganglion

The sensory ganglion of the metamere is a "Basic Stamp II" consisting of a 20 MHz PIC 16C57 microcontroller ( Parallax , Rocklin, CA). It is programmed through a proprietary Basic interpreter via a dedicated 9 pin RS232 serial connector from a portable IBM-compatible computer. Sensory ganglion has the function of collecting the sensory information from the legs of the metamere. For convenience, it also performs more general tasks such as decoding motion orders from the radio receiver and monitoring main manifold pressure. It then periodically sends a train of bytes to the brain, serially at 9600 Baud.

5.1.3. Brain

The brain is a "FlashLite" single board computer (SBC) consisting of a 8 MHz NEC V-25+ microcontroller with 512 kBytes RAM, a 256 kBytes flash memory disk, and two serial ports ( JK Microsystems , San Pablo, CA). It is very conveniently IBM-PC (Intel 8088/8086) compatible, and runs a brain.c program written in Borland Turbo C++, which is downloaded to it via a dedicated 9 pin RS232 serial connector. The brain interprets the sensory data as a function of the motion requests from the radio and sends off appropriate orders of the degree of opening of all the pneumatic valves.

5.1.4. Motor ganglia

The motor ganglia are "Mini Serial Servo Controllers", 8 MHz PIC 16C61 microcontrollers running a prewritten assembly language program ( Scott Edwards Electronics , Sierra Vista, AZ). They have the function of receiving orders serially from the brain at 9600 Baud and shaping them into an appropriate non-stop train of PWM pulses for each of the 12 servos.

     left leg sensors                    right leg sensors
+---+  +---+  +---+  +---+           +---+  +---+  +---+  +---+ 
|A/D|  |A/D|  |A/D|  |A/D|           |A/D|  |A/D|  |A/D|  |A/D| 
| F |  |FTA|  |CFA|  |TCA|           |TCA|  |CFA|  |FTA|  | F | 
+---+  +---+  +---+  +---+           +---+  +---+  +---+  +---+ 
  |      |      |      +---+       +---+      |      |      |
  |      |      +-------+  |       |  +-------+      |      |
  |      +-----------+  |  |       |  |  +-----------+      |
  +---------------+  |  |  |       |  |  |  +---------------+
                  |  |  |  | serial|  |  |  |
                  |  |  |  | 12 Hz |  |  |  |             +---+
 |                |  |  |  |       |  |  |  |             |amp|
+------+  PWM     |  |  |  |       |  |  |  |     serial  |---|
|RX    | 50 Hz +-------------------------------+  12 Hz   |A/D|
|progr.|-------|       sensory ganglion        |----------|PSI|
|   yaw|-------|     and upperbrain relay      |          +---+
| pitch|-------|        BS2 / PIC 16C57        |== RS232
+------+       +---------------=---------------+
                               | 13 bytes
                               | 9600 Baud
                               | 12 Hz
               |                               |
               |             brain             |
               |     FlashLite / NEC V-25+     |
               |                               |== RS232
                          | booster |
                               | 36 bytes
                               | 9600 Baud
                               | 12 Hz
              |                                 |
    +------------------+               +------------------+    
    |  motor ganglion  |               |  motor ganglion  |    
    |  SSC / PIC 16C61 |               |  SSC / PIC 16C61 |    
    +------------------+               +------------------+    
 +----+  |  |  |  |  +----+         +----+  |  |  |  |  +----+ 
 |    +--+  |  |  +--+    |   PWM   |    +--+  |  |  +--+    | 
 |    |    ++  ++    |    |  58 Hz  |    |    ++  ++    |    | 
 |    |    |    |    |    |         |    |    |    |    |    |
+--+ +--+ +--+ +--+ +--+ +--+     +--+ +--+ +--+ +--+ +--+ +--+
|FT| |ET| |FF| |EF| |FC| |EC|     |EC| |FC| |EF| |FF| |ET| |FT|
+--+ +--+ +--+ +--+ +--+ +--+     +--+ +--+ +--+ +--+ +--+ +--+
    left leg servovalves               right leg servovalves
Figure 9. Organization of the electronic modules of Autopod 0.33

5.1.5. Electrical power supply

Electrical power is provided by off-the-shelf Nickel-Cadmium batteries. Due to potential problems with electrical noise, separate power supplies were used for the nervous system electronics and the servomotor controlled valves. Good physical separation between power and signal wiring allowed a 75 MHz radio, several 20 MHz ganglia and a 8 MHz brain to work without interference on the same power rails, which was somewhat unexpected. Details of the power consumption are tabulated in Fig.10.

|       component      |  V  |   mW |  Ni-Cd battery  |
| 5V regulator         | 9.6 |   52 |                 |
| radio                | 5.0 |  145 |                 |
| pressure board & A/D | 9.6 |   73 |                 |
| potentiometers & A/D | 5.0 |  208 |                 |
| sensory ganglion     | 9.6 |   52 |                 |
| brain                | 5.0 | 1352 |                 |
| motor ganglia        | 9.6 |   94 |                 |
|                      |-----+------| high-discharge  |
| SUBTOTAL             | 9.6 | 1976 | 9.8V 500mAh .4C |
| left 6 servo-valves  | 4.8 |  300 | 4.8V 500mAh .1C |
| right 6 servo-valves | 4.8 |  300 | 4.8V 500mAh .1C |
| TOTAL                      | 2.5W |
Fig.10 Electrical power and energy budget

The brain is a TTL version and therefore consumes the bulk of the electric power. CMOS versions exist on the market and consume a tenth of that but are more expensive and less robust. The brain also has a "sleep" mode which was not used. Total drain on the nervous system battery is about 200 mA and drains a 500 mAh battery (~ 0.4 C) in 30 minutes, about the typical duration of a Search and Rescue run. A high-discharge type used in R/C racing model cars without the usual 0.3 C discharge rate limit were selected. Higher capacity batteries are available for a negligible increase of total weight.

The maximum drain caused by the servo-valves was measured by simultaneously activating all servos, back and forth. It is about 60 mA and well within capacities of standard 4.8 V 500 mAh batteries. This was a good surprise and a definite advantage of the hinged kink valves. Simultaneous operation of the 12 three- way proportional valves (600 mW) consumes less than one single on-off Clippard Minimatic solenoid valve (670 mW), which has much less air throughput. At any rate, with this system, electrical power for the valves is not a limiting factor.

5.2. Algorithms and software

One of the great ironies of the Autopod 0.33 Project was that designing, developing, testing components, and building the prototype robot took nearly one man-year, but writing a program to make it walk successfully following radio orders took just one week. This little fact has large consequences for the planning of future walking robots.

5.2.1. Mostly Proportional Control

Program brain.c is a proportional feedback system, with a few feedforward attributes. For example, when the leg is to be put down on the ground, it does not target some prespecified coxo-femoral angle, but opens all appropriate extensor valves full bore until the leg has established ground contact, and only then it adjusts valves to follow angle depending on the pitch joystick position.

The robot also possesses a potentially hardware saving reflex. If there is a loss of contact of both legs simultaneously, it means that the robot is very likely falling or about to do so. It does not have enough time to interpret what is actually going on and to take appropriate action. Therefore, the program activates each and every muscle actuator in anticipation of some shock, which can be assumed to be coming from anywhere, and this is enough to keep the robot on its feet.

Finally, Autopod has a couple of fail-safe systems. If the main manifold pressure drops below 50 PSI, or if radio contact is lost or too glitchy, the robot shuts itself down safely along a predetermined sequence of valve actions. Many animals walk successfully, but adherence to biological concepts and terminology, although advisable, was not mandatory.

The simple "mostly proportional control" program is enough to make the robot walk. Obviously more complex programs of PID or other adaptive control types can be experimented with, but the philosophy or robust walking machines should be to start simple and go complicated only if needed.

5.2.2. Serial communications

Unexpected problems developed with the serial communications between modules. First, the so-called "RS232 standard" is all but standard. The Clear-To-Send and other handshaking lines vary greatly from one manufacturer to another. Second, the NEC V25+ chip is not 8086/8088 compatible in the critical domain of serial ports registers. Some registers cannot even be read, as that chip is optimized for interrupts that were not used. Also, the "zapping" risk was real even for this TTL version during the experimentation, because some pins need to be brought directly to 5V or ground, some others need to be pulled up or down via 5 kn resistors, and microcontrollers and their manuals are generally not user-friendly. An oscilloscope was invaluable during that debugging phase.

Instead of using strobe lines or hardwired handshaking, in the end software handshaking was used, with reserved value 255 (11111111) acting as start byte for byte trains. The sensory ganglion acts as a Zeitgeber pulsing at 12.2 Hz, and the others follow. Measuring the times for each type of activity of the brain (Fig.11) shows the striking fact that the brain spends no less than 65% of its time sending or receiving data, at 9600 Baud, leaving only 40 ms for the "thinking" part of the brain cycle.

|                 period 82 ms  (12.2 Hz)                   |
|  receive   |   think   |           send            | wait |
|   17 ms    |   28 ms   |           37 ms           | 13 ms|
Fig.11 Time used for each type of brain activity in a cycle [return to top of document]

6. Performance

6.1. Walking results

The final result is that, given the stability and simplicity constraints of the wheels, Autopod 0.33 walks, untethered and semi-autonomously (radio-controlled), with the wheels in no way alleviating the loads on the legs. There is no doubt that the 6- legged untethered autonomous version 1.00 is possible. Swing time of a leg up in the air is about two seconds, and stance time on the ground several seconds, depending on propulsive load. Autopod 0.33 does not possess any data logging capacity, so the performance of the robot has to be evaluated by visual observation. For example, the two legs alternate for walking, but at one point the robot hit a large obstacle (on purpose) and the two legs decided to push together against it. This looked very life-like and as an "emergent property" but was just the result of the loose logic coupling between the two legs in the metamere algorithm. Simple visual observation can give many results and saves the complexity of a "black box".

Even with simple programs, Autopod 0.33 walks convincingly on a cemented or rocky mostly horizontal driveway. On a grassy and sometimes muddy real-world mild slope (Fig.12), some problems appear. First, Autopod 0.33 is unable to climb up a 10 thick grass slope. There are three interdependent reasons for that: 1) The robot has to pull up the slope a mass of about 50 lb (see section 6.3), all with only one leg. There would be from to three to five to do so in a 6-legged version. 2) The wheels pack grass or mud in front of them in miniature hills which locally have a slope well exceeding 10, like wheel chocks. The whole point of walking machines is that legs step over obstacles wheels have to roll over. 3) The horizontal swing moment of the legs is insufficient. The extensors and flexors of the coxa are undersized in Autopod 0.33 (Fig.7), and need to be increased in number to tackle slopes. The stronger the slope the more these actuators have to bear the weight of the whole robot (70% of the total weight for a 45 slope). A second problem is that the push-button leg contact sensor would not trigger consistently if the leg happened to sidesweep a tangle of grass or hit soft mud. As a result, and for the algorithm used, the robot could then "lock up". The push-button activation force is too high and it is not exposed enough to trigger in all conditions. This is relatively easy to correct.

Figure 12. Photo of Autopod 0.33 walking backwards down a 10 grassy
slope, in a light rain.

6.2. Energetic autonomy

The walking distance autonomy, which may be a wrong indicator of performance anyway, was not measured, but could be of the order of a city block, on a flat surface. Simple compressed air saving systems, such as having a flexible rod inside the McKibben actuators to occupy non-working and yet compressed air volume, or having the valves closer to the actuators to save dead space in tubing, may be used. More complex energy saving systems are possible by recycling the compressed air released directly to ambient air into intermediate stages of a compressor, in an action similar to compound and multiple expansion steam engines. Electrical systems autonomy is not a limiting factor (see section 5.1.5). Many so-called autonomous walking robots trail a fluid power and/or an electrical power supply tether, that Autopod was designed to avoid as much as possible given its intended mission.

6.3 Weights

Weights are all-important in walking machines, so a break- down of the weight of each of the components is tabulated in Figure 13, both for Autopod 0.33 and for an hypothetical 6-legged Autopod 1.00 using the same kind of components. Because of the geometric arrangement (cover photo and Fig.1), while walking or standing on one leg, each leg of Autopod 0.33 bears a weight of about 16 lb. In the alternating tripod gait of a 6-legged robot, each leg would have to bear on average a load of 17 lb (comparable to 16 above), or, in some extirpation case with only one leg up at a time, only 10 lb. It is therefore quite possible to double the tanks of 3000 PSI air storage. There is a new trend for 5000 PSI tanks in the aerospace industry, but calculations show them to have only a marginal increase of specific energy (good volumetric energy is not as relevant for walking machines), at more than twice the cost.

|    Autopod version (real & estimated)    | 0.33 | 1.00 |
|                components                |  lb  |  lb  |
| body frame & four low-pressure tanks     |  5   |  6   |
| servovalve board & low-pressure manifold |  2   |  6   |
| electronics board & Ni-Cd batteries      |  1.5 |  3.5 |
| legs with all actuators & sensors        |  6   | 18   |
| high-pressure cylinder, full, with valve | 14.5 | 14.5 |
| high-pressure to low-pressure regulator  |  3   |  3   |
| SUBTOTAL                                 |  32  | 51   |
| wheels & axle assembly                   |  18  |  0   |
| TOTAL                                    |  50  | 51   |
Fig.13 Component weight summary in Pounds

6.4. Costs

Autopod had to be inexpensive for 2 reasons. First, limited personal funds were available for the project. Second, "walking robots for everyone" were envisioned, rather than a one-of-kind expensive one. A quote of $25,000 for a light 3000 PSI compressor (used in jet fighters) that would have been ideal for the job, immediately closed the possibility of these expensive parts from the start. Component cost (Fig.14) was remarkably low for an untethered pneumatically powered walking robot.

|                      | Autopod 0.33 | Autopod 1.00 |
|      components      |       $      |   $ (est.)   |
| aluminum frame       |      30      |      45      |
| leg hardware         |      20      |      60      |
| nuts & bolts         |      20      |      40      |
| styroboard           |      20      |      20      |
| tubing               |      20      |      40      |
| fittings             |      20      |      30      |
| axle & wheels        |      74      |       0      |
| HP tanks             |     330      |     657      |
| regulator & HP hose  |     101      |     170      |
| LP tanks             |      10      |      20      |
| McKibben actuators   |      76      |     228      |
| servos               |     150      |     450      |
| kink valves          |      10      |      30      |
| potentiometers       |      60      |     180      |
| pressure gauge       |      25      |      25      |
| radio control        |     140      |     195      |
| sensory ganglion     |      52      |     208      |
| brain                |     218      |     654      |
| motor ganglia        |      96      |     288      |
| other electronics    |      40      |      40      |
| Ni-Cd batteries      |      50      |      90      |
| TOTAL                |    1562      |    3470      |
Fig.14 Component costs in Dollars

The development time was about 10 months for design and testing parts, and 2 months for the actual building of the robot. The building was delayed as much as possible so as not to crystallize the design too early.

The specific tooling costs comprised:

The operating costs are essentially the $15/tank refills (those are free to an University). No part needs frequent replacement, and this includes the kink valves.

There is a hidden cost of walking machines; they usually need to be brought to their location of use. By removing axle and wheels, Autopod 0.33 fits in a hatchback compact car, but a 6-legged version would require the use of a pick-up truck or the purchase of a trailer. [return to top of document]

7. Conclusion

7.1. Six-legged robot possibility

The main conclusion is that 50 lb untethered pneumatic walking machines for urban Search and Rescue are possible. Autopod 0.33 actually walks with actuators that are only 65% as powerful as the industry average, and with only 70% of a nominal 100 PSI operating pressure, so there is clearly room for improvement or significant payloads. A 6-legged version Autopod 1.00 is clearly possible.

The inability of Autopod 0.33 to climb a 10 slope can be solved in a 6-legged machine by increasing the size (here, the number) of actuators swinging the coxa. For a similar total weight, a six-legged version can also have up to 5 propulsive legs on the ground (as opposed to 1 in version 0.33), and no inert wheels to push above obstacles.

7.2. Underestimation of hardware design

Much of the study of walking coordination comes from biology, and therefore from the attempt at understanding principles of an already existing (living) object. The bulk of the research concentrates on the "smart" part, the nervous system, but the interaction between nervous system and physical entities like muscle, inertia, and scale is largely ignored. Of course there have been a few heroic attempts, including the work of the Hexapod Group at UIUC. It is easy to consider the hardware of the robot as an assembly of nuts and bolts to support an artificial nervous system, but in fact the hardware is equally important, perhaps even more so, than the nervous system. In other words, the controlling system is grafted onto the robot; not the opposite. Autopod 0.33 showed that the hardware takes about 50 times as much effort as the software, which is certainly contrary to generally shared notions about the problems of walking machines. It appears that a walking machine project without a clear definition of the hardware itself is simply doomed, no matter how smart the nervous system may be.

7.3. Brain organization

The biologically-correct metamere principle has promising features, but the slowness of serial communications between modules is disappointing. This slow pace is not a problem for Autopod 0.33, but could be for a three-metamere version. In fact, the UIUC Hexapod Group has had similar problems with a dead-end Ethernet avenue. Some microcontrollers use the similar I2C bus, which at least seems easy to use, but its real throughput and its timing limitations are not clear. Also, some microcontroller boards have two piggy-back processors, one to think, and one to handle communications, so slowness of communications is certainly not a new problem. As a result of all of the above, the best approach seems to be a single board computer (SBC) type with all analog-to-digital inputs on daughterboards with Direct Memory Access (DMA). A Pentium-class computer is certainly powerful enough to control a relatively slow-walking hexapod. The metameric principle and associated distributed architecture would be preserved only in software.

A great deal of work of the Hexapod Group at UIUC dealt with the control of the compliance of the actuators, muscle-like or not. The experience gained with Autopod 0.33 does not show this to be an issue as important as it looks in theory. Autopod 0.33 walks quite well in "hydraulic mode", so to speak. It also walks slowly enough that the low 12 Hz sampling rate of all sensors and commands is not a problem at all.

[return to top of document]
Back to Jan Cocatre-Zilgien's homepage