Section of Final Report: NASA ACRP NCC7-7, Bush Robots, Hans Moravec, Jesse Easudes

Project Summary

This is the final report for "Fractal branching ultra-dexterous robots," also known as "bush robots," a concept for robots of ultimate dexterity. A bush robot is a branched hierarchy of articulated limbs, starting from a macroscopically large trunk through successively smaller and more numerous branches, ultimately to microscopic twigs and nanoscale fingers. With its huge numbers of opposable fingers of all scales, a bush robot's dexterity would greatly exceed anything known. A fully realized bush robot remains a prospect for the distant future, be we were able to investigate on paper some aspects of bush robot geometry, structure, actuation, control and fabrication. We produced animations of limited bush robots. After briefly investigating and rejecting as beyond our means several approaches to building actuated models, we built two passive bush robot models, one a handmade passively articulated mechanical model with a 16:1 size ratio between its largest and smallest limbs, and the other a static model made by a solid printing process called stereolithography, with an 81:1 scale ratio, and 19,683 (3^9 ) smallest fingers.

We undertook this contract with the understanding that we were investigating ideas likely to require a half-century to mature, like the concept of rocket space travel when it was first seriously studied at the beginning of the twentieth century. Our experiences during the work did not disabuse us of that notion. Although some of the geometric structural issues in bush robot design were easy to resolve, we found no plausible approaches to building interesting bush robots in the near future. Available techniques fall far short of what is necessary to fabricate, actuate or effectively control such a thing. We did note an embryonic new manufacturing technique, that "prints" solid objects layer by layer, which might be extended to being able to

Enabling Technologies

Several key technologies must be much advanced to make possible practical, high-dexterity bush robots.

Computational power increased at least a millionfold in density . A bush robot, with thousands or millions of fingers or more, is much too complex to be controlled by a human being, so must have autonomous control. Even our simplest configuration algorithms running on good workstations take minutes to pose modest bush robots with fewer than a million fingers. Yet the natural frequency of motion of the smallest fingers in such a bush is in the kilohertz, about a million times faster than we could control. Higher degree bushes require disproportionally more control power, and more interesting grips or poses demand some degree of combinatorial search to find, boosting the requirements even further. Ideally computational units would be small enough to be distributed throughout the bush, down to into the individual twiglets at least a thousandfold smaller than existing chips.

Means to build mechanical structures over the entire range of scales from macroscopic to microscopic . In building our mechanical model, which uses conventional parts and a very modest scale factor, with limbs that range from one meter to five centimeters in length, we found it difficult to find properly sized components like tubes and balls and screws, and settled for 20% deviations from proper scaling. Shops that could manufacture our larger parts were hard-pressed to make the smaller ones. Our problems would have been be greatly magnified had we wished to actuate our model, or build it over a greater scale range. In fact, we abandoned an actuated approach using remote-control servos partly because we could not obtain them over an interestingly wide scale range. Ongoing research in micro electro-mechanical systems has promise, but cannot help us yet, for it is limited to mostly planar assemblies and works only at microscopic scale.
Our most promising fabrication used the "rapid prototyping" method called stereolithography. Using it, we were able to construct a solid static model of a high-degree bush with little more difficulty than it took to generate a graphic rendition. Stereolithography is a 3D printing process, that forms solid models as a computer-directed laser selectively cross-links and thus solidifies portions of a liquid polymer bath. With this approach we could construct assemblies with huge numbers of parts in exact sizes from 50 cm in down to the 50 micron resolution limit of the printer. We managed to produce bush robot models with 9 levels of three-way branching, far exceeding any other alternative. The printing hardware, in principle, might have produced even 10 or 11 levels of branching, but several tries at making a degree 10 bush failed, apparently due to memory limitations in the stereolithography machine's controlling computer. Very advanced future rapid prototyping machines, which could "print" in arbitrary materials, might allow working bush robots to be generated in this way. Once they exist, myriad-fingered bush robots, able to grasp and rapidly place tiny specks of matter into arbitrary configurations, could themselves become the ultimate rapid prototyping machines, able to assemble copies of themselves as well as virtually anything else.

Scalable high-power actuators . We briefly investigated the possibility of building an actuated bush robot model using standard remote-control servos, but rejected it both because the size range of available actuators was too narrow, because their one-axis actuation made three-dimensional motion very awkward, and because their power/weight ratio was insufficient for a self-supporting bush, especially in 3D actuators, where each joint required a minimum of three servos. We also investigated shape-memory wire. Though not very energy efficient, the wires have high power/weight ratio and can be scaled almost arbitrarily. One problem is that their cooling time grows about as the square of the scale for round wires, so fat wires are excessively slow. This can be countered by using ribbon geometries or parallel thin wires, but then complexity grows, especially for 3D joints. Also the wire operates in an essentially bistable mode, not suitable for continuously positionable joints, unless very complex schemes using multiple independently switched wires are used. Still, the simplicity and scalability of wire actuation is instructive. Actuators with similar geometry and strength, but allowing a continuous range of lengths, and compression as well as tension, would be ideal. We noted that a bush branch with six degrees of freedom could be constructed using just six linear actuators zig-zagged between a base and a rotated upper triangle in a "Stewart platform" configuration. Biological muscle, where myosin proteins "walk" actin filaments along them, provide an interesting model. Experimental linear actuators have been demonstrated, in which piezoelectric collars at the ends of a piezoelectric cylinder alternately tighten and release a central rod as the cylinder cycles from long to short, moving the rod in inchworm fashion in either direction depending on phasing. In principle such actuators could be made in a large range of sizes, if manufacturing technologies like those discussed in the last section were perfected. Electromagnetic or electrostatic analogs of the piezoelectric scheme are also conceivable, as are others based on mechanical effects otherwise driven, for instance optically.

Scalable high-density power distribution and storage . Distributing power to the larger branches of a bush robot could be done by straightforward conventional electrical cabling, but at higher levels, as the number of branches grows and their size shrinks, the wires would become microscopic and very numerous while needing to tolerate ever higher-frequency flexing, multiplying the probable failure rate. Connections, flexing and bulk might be reduced if power were stored within individual segments, and replenished in a bulk way, for instance by placing portions of the bush in a radio frequency field, with segments behaving like antennas to couple to the radiated power. Local storage might be accomplished by very high performance capacitors, made of microscopically precise conductor/dielectric structures.

High-strength materials . Scalable computation, construction actuation and power, would be enhanced by the perfection of materials technology. In particular, carbon nannotubes, with nanometer scale, strength to weight ratio up to 1,000 time steel's, and versatile electrical properties, would be suitable for almost every aspect of a bush robot assembly, and easily provide the light weight and high power needed for self-contained actuation. Conventional structures and actuators, let alone power sources, have a hard time lifting their own weight in a manipulator. Most dexterous robot hands, and even the original biological human hand, are operated via tendons actuated at a distance by heavy muscles further up the arm. Such displacement of actuation would be an awkward complication in a bush robot, where fingers themselves have fingers.


Fully realized bush robots are so far in advance of available technologies that there is little urgency to pursue their theoretical development now. Doing so is an amusing diversion, and may by chance lead to interesting insights, but is no more likely to lead to practical devices for many decades than other lines of undirected research.

On the other hand, our investigation so far illuminates a systematic deficiency in present day technologies, having to do with scale. Traditional manufacturing methods evolved from manual fabrication, and are most effective at the scale of the human hand. They encompass a vast range of structures and materials, but become impractical at sizes below human finger precision and visual acuity. They also are complexity-limited, since humans can do fewer than a million modification steps in a month, even at the rate of one per second. Larger constructions are possible using repeated applications of human-scale work and more humans, as in bricklaying, and more recently through the amplification of human size and strength by powered machinery. There are no good means for applying traditional construction methods much below human scale.

Recently, the entirely new fabrication approach of photoreduced lithography used to make integrated circuits has allowed structures a thousand times smaller than the human lower limit to be built, and advances in microscopic projection have pushed the minimum scale down another thousandfold. But the common lithographic methods are limited to near-planar surfaces and certain materials, and are economical only for small objects.

The brand-new method of stereolithography and other "rapid prototyping" techniques, that form large objects layer by microscopic layer, are limited neither by human scales, nor the flatness and size limitations of conventional planar lithography. The stereolithography process is very much like bricklaying, in that a large number of identical standard pieces are assembled to make the final structure. Unlike human bricklaying, however, the piece size is microscopic, about 0.05 mm in common machines, and the number of parts that can be assembled ranges to 10^12. By comparison, a human being, moving one part per second, could assemble only about 10^7 units in a year of work. The smaller the brick size and the larger the number of bricks, the greater the range and functionality of objects that can be constructed. Some rapid prototyping systems are able to put down a variety of materials, analogous to color printers which can deposit a variety of inks. None yet has the resolution or material range to make adequate moving or electronic parts. Experimental silicon micro-electro-mechanical methods are able to make some working parts of very limited height. It is plausible that the state of the art can be improved, perhaps by marrying and extending stereolithography and micro-electro-mechanics, to the point where, for instance, a working mechanical or electronic watch could simply be "printed." It will require only small advances beyond that stage to print working bush robots of modest degree, perhaps using physically-based scaling laws to derive each level's structure, actuation, power and control from a base design.

Ironically, once there are working bush robots, approaches resembling manual manufacturing methods might undergo a revival. This is because bush robots have "hands" at all scales, from the macroscopic to the microscopic, possibly to the nanoscale. Equally important, the number of bush robot hands, as well as their speed, increases as the scale shrinks, matching the growth in the number of components found at increasingly small scales. Bush robots thus overcome the scale and quantity limitations of human manufacture, while retaining the flexibility, and will be able to assemble objects more flexibly, diversely and efficiently than simple layer-by-layer lithography. With sufficiently intelligent controls, bush robots should also be able to make repairs and alterations in existing structures, including biological ones, which lithography simply cannot do.

Summary Conclusion

Because of fabrication, scale and control issues, bush robots do not seem a practical approach for robotic manipulation in the near term. Those needing robot arms are better off choosing approaches that are attuned to human scale, manufacturability and are ease of control, all suggesting a parsimony of complexity.

Bush robots are for a future time when structural, mechanical and control complexity costs little. The arrival of such a time can be hastened by encouraging the development of uniform automatic manufacturing methods that can build to the size and three-dimensionality of stereolithography, while exceeding the precise shape control and materials flexibility of experimental integrated-circuit-based micro-electro-mechanical methods. The latter research path is sure to provide many rewards long before it enables the construction of bush robots. But when effective bush robots are constructed, they may provide superior and vastly more flexible fabrication means that eclipse the effectiveness of their uniform lithographic parents.

Manually constructed passive articulated bush robot model, branching factor 3, taken to degree 6, 2.1 meters tall: Because of cost limitations, only one of the three major branches was constructed, and only one of the three second level subtrees was filled to degree 6, the other two subtrees are populated to degree 5. Because of limitations in available parts, different materials were used for different level ball joints, and various levels deviate in scale by about 10% from perfect proportions. There are a total of 135 end fingers, 81 at level 6 and 54 at level 5.

Stereolithographically printed bush robot model, branching factor 3, taken to degree 9: All parts are in near-perfect proportion, and all nine levels are complete. There are 3^9 or 19,683 end fingers. The model is 11cm tall.