Naval EOD Actuators and Legged Locomotion workshop Notes from M. Cutkosky Chris DeBolt at Naval Explosive Ordnance Disposal (EOD) program opened meeting by showing the EOD recruiting video. (Get hazardous duty pay and work on things that hopefully don't go bang in your face.) So they would like to see ordnance disposal robots in the field. Before the presentations began, Bob Full talked with me about consulting with Edge Innovations Inc. on insects for a giant insects movie. Lots of parts to look at. Lots of sizing issues. Can we get some after-the-fact contacts? Who are their artisans that make the insect parts? Think about how they actuate what they built... Are they willing to talk about it after the movie comes out? ======================================================== Presentations Tues May 17: Ronald Pelrine (SRI) Can use a variety of elastomers for the contractile polymer actuators. Polyurethane (Deerfield PT6100S), Silicone,etc. Fabrication uses polymer thin film techniques. Process involves solvent casting, dip coating... Spin coating - gives highes quality, thinnest films. Looking at if for multilayer consructions. Electrodes - compliant and conductive and pateernable. colliodal carbon, carbon nano fibrils, gold (lower strain <5%), organic salts. conventional depostion spray, cast, sputter, casting with selective wetting, spin coating. Need intimate contact between conductor and polymer. Energy density, efficiency, working strain, working stress or pressure, mechanical impedance. Other concerns - operating voltage, cost... "Actuator Comparisons" Nice slide - can we get a copy of this to include with attribution for our book on Human and Machine Haptics? Ans: yes. Need to check with Jorge whether this really has all the info. we'd want. Challenges: In-situ microfabrication of multilayer designs. Scale up to larger actuators - film quality, fault tolerant designs with self-healing electrodes. better materials with lower voltages for high force Couple muscle to over-center mechanism to decrease effective impedance for initial displacements. Voltages today like 200-5000 volts depending on film thickness. But 1cm^3 voltage converters are available these days. Safety - very low current. (A thought... Anybody ever used piezo electrics as a voltage conversion device? Or maybe for thermo -> high voltage.) Applications: Japanese micromachine (MITI) robot for pipe inspection. Who? We should track down refs on this.... Sonar MEMS noise supressor Flapping Wing propulsion for DARPA (just started). (Does Fritz know about this?) Pumps, valves, printing heads. Why use it for legs? many shapes: Rolls, tubes, biomorphs. Can also think to configure as clutch, linear-t-rotary device. Speeds to ~1Khz feasible. depends much on form factor. 100Hz for std. tube things. Similar density to electromagnetic motors. Scales better, they think, to small robots than EM motors. Large displacements reduce need for transmission, improved back-driveability. Looking at 6-leg robot at SRI. Need force sensing - can use these devices as capacitive or resistive sensors while actuating. Programmable compliance -- can put multiple degrees of freedom on the same substrate & film (e.g., for increasing dynamic range) enhanced output at resonance (not possible with other actuators that are limited by saturation). Direct insect leg actuation feasible. Adequate range of motion? Cascade and bundle? These devices oughta be more stealthy that std. motor-driven stuff. What are maximum tensile stresses for these actuators? Working on a leg for smaller robot. 20-33gm force +/- 35deg motion. 3DOF. Maybe we should show these guys Mike Binnard's spreadhsheet of requirements and use it as a basis to think about appropriate sizing & geometry? Teresa McMullen (our ONR officer) - why shoebox size robot? Where did that idea come from? ANS: Talked with Navy and that was what they suggested. Teresa -how many of these actuators can also be used as sensors? Ans: most. Action item: get Jorge, Mike, Beth to come with me to visit SRI and learn more about details of electrostrictive polymer fabrication at micro and MEMS scales. Also get the SRI folks to visit CDR and RPL. I think we should plan this summer to either embed one of their actuators or else we fabricate one in-situ as part of our fabrication processes. Note: I have a copy of Ron's paper from IEEE ICRA'98. This is a good overview to the technology. ====== David Wilson & Mehran (?) Shahinpoor (?). Biomimetic Products Inc. N.M. Actuation with small applied voltages. Ionic membrane. 1-50Hz microactuator. Strain ~10K microstrain. Milivolt output when used as a sensor. Efficiency up to 96% they say (?) Question: Efficiency is function of material thickness? Ans. mumble. Oscillating voltage input and load characterization. Thin film with charge -> internal field 10^10 v/M. Something mumble ion dispersion in the film => strain. (Wilson, the presenter, was not the expert on how it works). Some voltage load/frequency plots. But no comprehensive characterization. Teresa - how does response depend on thickness and length? What's the time constant as a function of length scaling? Ans: We don't know yet. Response evidently depends on time profile of applied voltage (due to time constant and viscoelastic effects). The stuff needs to stay wet. Made from chemical plating ion exchange membrane polymers with platinum. Claim the result is low impedance and not too hard to fabricate. Beam theory model for sensor appliation; end load => strain => charge. Signals on the order of millivolts for mm of displacement of a biomorph. A web site exists at http://www.biomimetic.com ============================ Marc Olivier SARCOS High performance wobble motors. (with Jacobsen, Simonson) This project has been going on for a long time, on the back burner at SARCOS. Also doing some work on sensors (accel &) for embedding into structures. Sensor to give information about impending failure. For DARPA. Wobble motor - cyl. rotor inside cyl. stator cavity. Rotator slowly rotates about own axis as it wobbles like a planetary gear. Stator coils energized sequentially (like stepper motor). Electrostatic or electromagnet. Variable reluctance example shown. Lately trying to build splines and some switching electronics etc. on cylindrical surfaces. Flexible coupling from rotor to outside world. Made it of a catheter. Flex in bending, stiff in torsion. 250 micron rotor diameter. Micro EDM method for electrostatic. 400rps. Ran for ~1 year. Put a 25 micron drill on it and drilled a hole in human hair. Now think about applications: Example, consider the SARCOS Dexterous Manipulator with hydraulics. Actuators: 7Nm finger pinch. 400 deg/sec rotation... Hmmm... wobble still not good enough. Also did disk whirl and other kinds of wobblers. Swash-plate-like things. Friction transmission mechanism limits torque. Can add gear teeth? Looking at better sensors and commutation schemes with ONR funding. Various big whirl motors built from 3 inch down to 3mm. Pager motor size. Coupling by U joints, flexible shafts (catheter tube), planetary gear... Looking at various in-situ sensing for commutation control: capacitive, hall effect, coils. Measure the changes in reluctance (varying gap when rotor occupies different positions in the cylindrical cavity). Hmmm.. This seems similar in principle to those harmonic drive spline stepper motors that came out some years ago. What was the brandname of those devices? ====================== Integrated Force Arrays Scott Goodwin-Johansson (MCNC) (Tom Kenny may be familiar with this stuff...) Started under NIST & DARPA. Now with ONR $$$. Polymer hollow rectangle with electrodes. Apply voltage ==> squeezes like a dogbone or bowtie shape. Make an array to get a force cell array. Polyamide patterned arrays. 2 microns tall, 1.7 micron cell pitch. Release layer, polyamide, gold/chrome conductive layer evaporated on. O2 plasma etch to remove all exposed organics. Finally put in HF acid to get rid of oxide leaving you with a freestanding mesh that you can drape or suspend. Photos.. looks like pretty crisp feature definition. Forces ~15ergs/mm^3, up to 20K contractions/second in arrays. Lifetimes > 10^8 cycles. Achieved > 12 dynes (10^-5N), stress ~3000dynes/mm^2 Hmmm... this is not as good as the SRI actuators. It may be that they are limited by air dielectric and by the bowtie mechanics (less useful force at beginning of stroke). Application: read write head positioning on disks (augment voice coil macro actuator with fine actuation) prosthetic and robotic devices steerable catheters Can use a sensor as well as actuator (of course) and as electrical power generator. Nonlinear response. Force starts roughly squared or exp. (with decrease in gap) and then snaps shut when about 1/3 closed. Showed a Strange theory vs exp. plot - not nearly enough data to tell if it's even close to the theory. When you apply these things in practice, non-uniform stresses over the sheet tend to result in some cells being over expanded and therefore not as effective. Can build mechanical guides, battens into the arrays to counteract local buckling or over stretching. Video of array of 3 columns x 100+ rows in a piece of tape. It contorts itself and writhes. Very fast. Looks like insect larvae. Looks like maybe 10% strain maximum (?). Now looking at polyamide arrays in a stiff frame. (Fabricate using MEMS technology). With plated gold bumps for connections. Going for taller plates => high aspect ratio structures. Try making a mold using LIGA and then mass produce. Use conformal CVD depostion to get metallization etc. ============================ Dan Urry (U Minn) Elastic protein-based polymer actuators. Hydrophobic association transition in model proteins. Phase separation on raising the temperature (Polymer becomes insoluble). Result is associated with folding of protein. Basic version - about 5 minutes to achieve 50% contraction! It's diffusion limited, but seems like hard to turn it into a practical actuator. Long talk... Jet lag strikes. ============================ Cutkosky - We had of interested questions for our project. Note to self: Add a couple more SDM concept slides before shipping the set to Teresa McMullen and Chris Debolt. Process examples and what can be built (take from Part I of E611 talk) - this audience doesn't know anything about layered mf'g. Dinner at Captain Billy's crab house. Dan Kodischek has ideas about differential geometry concepts and finding invariants that could be applied to creating 'biomimetic' primitives (as opposed to the simple geometric shapes we've been using) from which to build robots. An intriguing notion, after a couple of beers... ============================ Wed May 18 Bob Full It appears that spring-loaded inverted pendulum works as universal principle of locomotion in virtually all legged locomotion. Morphology doesn't make much difference for straight-ahead locomotion. So what do we attribute the huge changes in morphology to? Think about stability (e.g., lateral). "Sprawled posture" animals - timing is different. Front leg slows body down intially, etc. Result is a self-stabilizing system. Example dynamic model with body as rectangular prism suported by spring legs. (Looks like the 2D simulation was done in Working Model.) Drive legs according to known pattern. It converges to a preferred speed. And it recovers from perturbations, as long as you have some basic joint-space stiffness and damping. The morphology plays a role here - kinematics are such that they helps stabilize over the range of leg configurations. (I bet this would be analogous to a stable grasp in dexterous manipulation - be interesting to look at the eigenvalues of the stiffness matrix...). Quite different rates of recovery for fore/aft versus lateral and rotational perturbation. (So what are the corresonding stiffness terms? -- of course the joint angles are large enough that need to consider how matrix varies over the working trajectories of the legs). Plotting various velocity components for many perturbations. First stride very rapidly recovers about 80% toward the steady pattern, then they (th velocity vectors) move along a predictable recovery "surface" to converge to the steady pattern. Full showed a velocity field return map. Note that if you try to do linear control to pull each DOF back, it actually takes longer to fully recover. It's better to take a characteristic return path that quickly gets you to the recovery surface. Dan Kodischek collaborating on this (aha - sounds like Dan). I wonder what Rob, Reza, Kazerooni think of this? It sounds reasonable. Other findings - centipede uses same basic strategy as cockroach, with clusters of legs. Same self-stabilizing pattern emerges. Tested ideas with cockroach walking on random fractal surface. Plots of body roll, pitch, yaw look pretty random... But the gait is basically the same as on flat terrain. And EMG plots look about the same for flat terrain and fractal surface. Cockroaches don't even slow down. They bump and jostle a lot (not graceful, and not like a person stepping from boulder to boulder at the sea shore). Maneuverability Numbers -- Did plot of force needed to accomplish a turn as a function of foot placement w.r.t. coordinates at center of mass of body. Shows the middle legs are not very useful (?) for turning. But I'm not clear on what the mathematical definition of the maneuverability number is... Extensions to consider gripping and slipping (all work to date assumes no slip). Looking also at crabs with dactyls (points) on feet. Dactyls work very well in sand (hmmm, not obvious). Studies of Gecko climbing on see-thru belt which can be horizontal or vertical. Look at effects of foot adhesion. The toes peel when the foot is ready to lift. Bands of very fine hairs in lamellae on the feet. Vanderwall's forces exploited? Would like to measure force of single hair with atomic force microscope. Are the surfaces self cleaning? Neuromechanical model looks a bit like Reza's. Control embedded in nice dyamically stable compliant system. Alterations for trajectory control are layered atop robust feedforward program. Having effective feet helps. Q&A session Q: when cockroach on rough terrain is contact always at the foot? Ans: no, sometimes on other parts of leg and even dragging body. Can even get some legs stabilizing the body (like gymnasts do with their arms when airborne). So 3D dynamics get complicated. Q: Your EMG plot (running on fractal vs smooth) is just one set of muscles in one joint. And not a set of muscles that would be most likely to change. What happens if you look at other muscles? Ans: good question. ======================================= Dan Koditschek (U Mich) - starting collaboration with Full. Robot ---- Environment Dan is interested in modeling the with nonlinear differential geometry ideas. Return maps, potential functions. Create "energy landscapes" and return maps on them to think about behaviour and convergence to goal states. Looking mostly at 2DOF archetypes. Need a new template when there is a change in topology. Brief review of his previous juggling and similar work in robotics. (Same notion of convergence to stable gait, return maps.) Vertical template embedded in 3DOF paddle ball task, juggling 2 balls. Use it to overcome obstacles (where do you have to be in the map to clear this obstacle and be in convergent space when you reach the other side). Unimodal return map (Raibert). Lets one see effect of adding energy: "Local computation gives global convergence." No Lyapunov function available. Invariant submanifold corresponding to the vertical template... Dan, what is actually going on here? This is becoming kind of hard to follow. Basic idea seems to be that as you go from very basic one dimensional juggler to two and three dimensional problems, you can still exploit the same basic vertical "template" to maintain convergence to a steady periodic juggling gait. (Similar to findings by Raibert et al when going from planar to 3D hopper and then two-legged machines). Mentioned work with dynamicist John Guckenheimer at Cornell. Now consider a 2DOF hopper type system. Is non-integrable. You just have a return map. Mappings between positions and velocities at bottom and top of cycle are crucial, but unsolvable. Cites idea from biomechanics literature McMahon and Alexander. "SLIP" Animals as tunable springs. Fit this into the Lagrangian field to see how well the model works as far as converging to expected minima. "SLIP" method from biomechanics - center of mass still goes like it's on a pogo stick as you go to bipedal running. (?) Ankle, Knee, Hip model. Again, embedding the map as attracting invariant submanifold works surprisingly well to collapse the complexity of the system. The details of anchoring are being worked out. Simulation results match measurements quite well. Show that differences are within exp. error so that "the fit does not refute the map hypothesis." I guess if we're to really get into this I need to re-read Dan's old juggling papers that I looked at some years ago with Brent Gillespie. ============================ Randall Beer (Case Western) Studying stick insect gait and dynamics. Think of legs cycling between front and rear "finish lines." The line locations vary dynamically as a function of what other legs are doing. Simulation data plotted as phase maps for dynamic model of 6 leg robot & distributed controller. Things work much of the time but there are strange (and previously unexplained) divergences for particular combinations of parameters. So now they too are looking at return maps. Except that these ones appear not to include full dynamics -- they are kinematic. Phase/velocity plots => regions in space for convergence. Have a basic clock cycle associated with gait. If leg is at state "A" on clock tick "i" where does it go at "i+1" ? Can write out, analytically, the expressions for what the return maps look like. To go for multiple cycles, can do maps iteratively till get an asymptotic return map (convergent). Can also extract phase information. In particular, can establish theoretical boundaries for phase regions. Can also do something like a probability density function. Can ask how likely, for given range of initial conditions, that you will converge to a certain set of phase relationships. ===================== Roy Ritzman (Case Western) "Neurobiology colloration with robotics - is useful to make you think about the mechanism (dynamics) that motor activity is associated with." Exploit nice one-to-one correspondence between motor activity neuron and muscle activity in insects (muscles have one activating & one inhibitory). Like Full, working on cockroaches. Powerful push movement, helped by well suited kinematic posture, in rear legs toward 2nd half of stroke delivers most of forward motion. There is evidence of recruitment of faster muscles for parts of stroke where that is needed. Lots of careful cutting into and inserting electrodes into nerve fibers of roaches. Sounds difficult to do. Example: if one cuts neuron to muscle between coxa and body (?) what happens? Insect gets strange lateral waddle shown on video. Looked at roaches climbing barrier about 1/2 body height or more. (More substantial barrier that Full's fractical surface.) This slows them down. Cockroach measures barrier with antennae, then front legs go right to top surface of barrier. Middle legs help change attitude of body. Push with rear feet. Leg searches for foothold, push again. Clamber over. They are studying this process in detail. ========================== Roger Quinn, Case Western. Working with Roy. Cockroach inspired insect robot about 2 feet long. Stable posture controller. "It walks, but not very well yet." (24 DOF in total: rear legs 3DOF, middle 4 DOF, front 5 DOF). Has typical cockroach stance. Can lift 30lb weight in addition to body weight and it can jump. Teresa - legs are kinematically similar. Dynamically? Ans: no. Not practical with pneumatic cylinders. Weight: frame = 7 lbs, total weight with valves, hoses, cylinders is 30lbs. Small compressor with model airplane engine if you want to cut the umbilical cord (very noisy). Uses a lot of air - not efficient. Can easily jump. How to keep it stable? Worked first on posture control since bio research suggests that's essential as a basis. It's more than local reflex interaction. It is orchestration and tuning of relfexes according to central desired behavior. Only sensors used so far are potentiometers on the joints. They do "virtual model control" for body, adapted from Raibert and Pratt. Now mounting seminconductor strain gages on the legs. Will initially use them as they did on older electric motor driven robot. Control: Input = desired body motion. Find desired forces and torques on body to get that motion. Find torques at each leg joint for desired body forces and torques and a given center of pressure. Use optimization to find solution consistent with these goals that gives good load distribution and overall joint torque minimization. (min max?, sum squared?). Ans: sum squared. Is this what we find in biology - more or less, but we're not entirely sure. Also taking friction limits at the feet into account in the program. Previous 6 legged with electric motors could walk over a grating with foot placement driven by force sensing. Briefly saw video. Valves PMW controlled at 50Hz (Buzzzzz). The valves have 200Hz bandwidth. They are derived from dot matrix printer heads. Lots of plastic tubing everywhere. The cylinders are metal. Bimba? Or maybe something slicker like Airpel? I think Bimba, he mentioned friction... (Confirmed during lunch). Says seal stiction not a problem when you use PMW control of valves :-) And apart from stick/slip, the friction actually helps to stabilize. (hmmmm). It's stable and compliant. And unfortunately, still has the buzzy, soft and not very responsive character that I associate with most pneumatic robots. Definitely not smooth or efficient looking. Free motion trajectory following has that characteristic tremulous look. (Gets better with adapation.) Looks better if you turn off the sound :-) Postural control is pretty bouncy and stable. Teresa - why not use gyros to get rate & acceleration information? Ans: not found in biology. Well, actually there is vestibular system, but evidence is that proprioception dominates. Need to augment posture controller with locomotion controller and appropriate reflexes for encountering rough terrain features. Looking at an adapative pattern generator for the locomotion. Desired vs actual foot trajectory. If you damage one leg slighly on cockroach, it limps and then controller adapts to mostly remove the limp. Leg learns joint motions (set point trajectory) needed to actually follow the model path. Virtual joint angle, or attractor idea. Demo, grab front legs and pull toward you. Uses posture control for stance and swing pattern (stick insect) to drive middle and rear legs. If you pull harder or slower it reacts appropriately. Controller is done on one Pentium class machine (50Hz valve PMW rate means you don't have to go all that fast.) If you try to do just joint space proportional control it goes unstable right away. Huge flapping, buzzing motions of limbs. Need large feedforward terms and posture stabilization. Need to individually calibrate the valves. These are obtained from some manufacturer. Working with Sasha Zill - make 3D realistic model of cockroach front leg using Zill's data. 1mm = 1inch scale. Put accurate dimensions, insertion points in AutoCad model. How was this model made? Machined aluminum plates 1/8" thick. Little springs to passively load the joints. Simplified, but approximately matches silhoutte. Using MKibben actuators (non linear - looks like they are just surgical rubber tubing inside protective braid sleeves of the kind used in wire harnesses). 14 tensile actuator groups. 5 tarsal segments so far. Strain gages. Aluminum and plasic exoskeleton. Also working with Zill to do FEM analysis of cockroach exoskeleton. USing Algor FEA package and meshing to get stresses, strains, deflections. Use results to know where to put strain gages etc. ========================================== Sasha Zill (Marshall Univ.) Neurobiology of sensors in arthropods. Coxa-Trochanter-Femur-Tibia-Tarsus (from upper limb to foot). Trochanter is short elbow-like thing. The tarsus is quite flexible. Largest muscles act on Tarsus - either inside body cavity or intrinsic to Coxa. So lots of force sensors at femur and tarsus. Receptors clearly seem to trigger strong extensor activity (correlation shown in studies). Neurons go into exoskeleton "caps." These are oval craters with domes. Measure strains in skeleton. Maximum sensitivity is for compression along minor axis of the ellipse. They tend to occur in clusters, some redundant, some with different orientation. They are relatively fast acting, but do have some DC response. As the animals grow, the caps grow and increase in number too. And they become sensitive to higher force ranges. Added white noise stimuli and low-pass filter to some rolloff frequency and measured receptor response. Result is linear best-fit bode plot. Looks like they are velocity sensitive. 1st order system? At any rate, sensor gain shows approx linear velocity dependence. Bandlimited white noise stimulus to receptors also leads to strong reflex muscle activations. (Apply force on tarsus.) Get a positive feedback: foot contact => sronger muscle response => stronger receptor feedback to quickly ramp up to some maximum force. Has detailed STL-like model of trochanter inside and outside. Done like CAT scan with sections and re-assembly. Wants to kow about possibility for rapid prototyping via stereolithography, etc. They also have a meshed FEM of same. 20 micron resolution mesh. Added in appropriate elastic modulus, etc. Major meshing effort! Von Mises strains correlate nicely with areas in which skeleton is darker and thicker, stronger. DiCaprio modeling responses of force receptors - correlates rather well with what has been measured. Tarsus - has no musculature. Muscle is up near top of leg with tendon to tip. Tendon pulls down & engages claw. What disengages? Passive? Engages immediately after contact (based on sensing contact?). They are held in up position when leg swinging. Rich in sense organs force, chemicals, joint angle. Little sensor activity during swing. Yes, it is elastic -- passive mechanism disengages it.