To wrap up the discussion of technical topics, let's go over a collection of practical approaches that should help you create useful and reliable mechanical designs for small- to medium-sized projects - no matter if you are using CNC machining, 3D printing, or still carving stuff in stone. The failure to get even the most rudimentary understanding of these topics is one of the cardinal sins of hobbyist makers - so don't let this happen to you.
The mathematics of material engineering are fairly obtuse, and the finite element modeling software that can be used to model the dynamic behavior of real-world part geometries is prohibitively expensive. Nevertheless, if you are designing small- to medium-scale components and can afford to learn from mistakes every now and then, there is a couple of simple rules of thumb you can follow to get great results without resorting to any of that.
(You can also find quite a few good engineering tips on this page.)
Perhaps the most profound cognitive dissonance you can experience during your first adventures in DIY manufacturing is making the first part that happens to be long, thin, and flat. Let's say you grabbed IE-3075 and cast a piece that is 2 mm thick, 1 cm wide, and 15 cm long; you will be probably impressed by how difficult it is to break this sample - but also shocked by how easily it deflects. It just doesn't seem to make any sense: the datasheet says that the material is ought to be 50% more rigid than the plastic used to make Lego bricks, and several times less flexible than the commodity plastics used in everything from toaster ovens to computer cases. But your toaster doesn't flex nearly as much, right?
To understand what's going on, consider what would happen if you applied the same bending force to the side of the part, rather than the top:
If you do that, you won't be able to flex the part even the slightest bit. The cross-section of the material is still the same - 2 x 10 mm - but all of sudden, it is a lot more willing to resist your efforts.
That's because flexing the material amounts to stretching the polymer chains farther apart on one side of the part, and squeezing them close together on the other end. If the material is thin, the effect is relatively modest, and you don't need to invest a lot of energy to make it happen. For thicker segments, the displacement is a lot more dramatic, however:
The effect is pretty significant. Simplifying slightly, the rigidity of the part is proportional to the width of the sample - but to the cube of its thickness.
This behavior brings us to the manufacturing trick that is being used in almost every item made out of plastic or sheet metal - including toasters and Lego bricks. In essence, you can greatly improve rigidity by increasing the maximum span of the part in the direction subjected to bending loads - without adding any significant bulk. The most common approach is the addition of a couple properly oriented ribs:
The average thickness of the part - and thus its weight or the amount of plastic used - hasn't changed significantly. But because each of these ribs contributes to rigidity in proportion to the cube of their height, their impact is more significant than it may seem.
The following picture is a good example of the use of ribs to control the rigidity of a thin-walled chassis of a medium-size robot. Note how the ribs are routed around any locations where additional clearance is required, and not interfere with the mechanical design. In fact, they probably improve the overall aesthetics:
It is also possible to achieve the same effect without having to introduce any "standalone" ribs; for example, if the part uses curved surfaces or happens to have outer walls, this can provide a comparable degree of reinforcement:
Properly reinforced thin-walled parts weigh less, cost less, provide more clearance for other components, exhibit much less shrinkage - and often simply look better. Because of this, try to learn and use these approaches as often as possible. Stick to 1 to 1.5 mm as a baseline wall thickness for all the larger parts - especially for chassis components, covers, etc - and just sprinkle them liberally with ribs, gussets, and other features of this sort. A handful of ribs extending 2-4 mm from the surface of the part is usually all you need to make it work.
So far, we've talked about the need to reinforce thin, long, planar surfaces; but there is one more situation where a small amount of extra plastic routinely saves the day.
From the discussion of casting resins earlier in this guide, you may recall that the resistance of these materials to non-uniform loads (e.g., tearing) is dramatically lower than the ability to withstand uniformly distributed stress. For this reason, you should pay close attention to sharp corners, jagged edges, holes, and similar features in locations subjected to significant loads - and incorporate simple reinforcements to ensure that the stress isn't concentrated in any single spot.
For example, when designing a coupling for an axle that may be subject to strong radial forces and shock, adding a fillet or a couple of load-bearing gussets can save the day; all other things being equal, the "vanilla" approach shown on the left is much more likely to fail:
Given the negligible cost of these features, it typically makes no sense to spend too much time trying to figure out in which exact spots they happen to be absolutely necessary. Whenever it is clear that a localized stress may occur, throwing in a rib, a gusset, or a fillet is simply a good plan.
Another problem that a novice designed may bump into is the question of fitting tolerances: when one part needs to mate with another, what should their respective dimensions be to form the desired type of fit?
The correct answer depends on a number of factors, including the accuracy of your CNC mill, or the operating conditions of the assembly (temperature, risk of contamination, running speed). Nevertheless, when working with a reasonably accurate milling machine and dealing with normal circumstances, the following rules of thumb tend to work pretty well for rigid plastics:
Snug fit: if you want the parts to mesh tightly - for example, to be glued or fastened together - make the opening exactly the same size as the mating part.
Interference fit: if you want the parts to hold together without any additional fastening, you can undersize the opening by about 0.01 to 0.02 mm (i.e., offset the profile curve by 0.005 to 0.01 mm). Note that you will need to use force to make them fit, so this may be not an option for fragile parts or unusually brittle resins.
Low-friction meshing: to allow a round element to rotate with minimal friction, it's usually sufficient to leave a clearance of around 0.01 to 0.02 mm. Note that for high radial loads, ball bearings may be preferable to sleeve couplings.
Self-tapping holes for machine screws: measure, look up, or calculate the minor diameter of the screw (the outer diameter minus the depth of the thread). Use this as a diameter for the hole. A drop of WD-40 will help drive the screw in for the first time.
When in doubt, be quick to experiment: it takes just a couple minutes to machine a rectangular or cylindrical test piece and play with it. Save yourself the time and frustration of having to debug issues in more complex parts.
Gears are an extremely versatile tool that will come handy in all sorts of mechanical assemblies. Depending on your needs, they can reduce or increase the speed of movement, adjust torque, or change the direction, axis, and type of motion - all that with high precision, excellent efficiency (95%+), and great load-bearing capabilities. Alas, unfortunately for hobbyists, much of the freely available information about the geometry of common types of gears is misleading or incomplete - and even the articles posted on Wikipedia are littered with inaccurate images or animations such as this.
The key property of a well-designed set of gears is that they transfer motion almost exclusively through a rolling action, with very little sliding (and thus friction) involved; and that they maintain constant velocity and a constant angle of contact through the process. If you violate these principles, your gears will become inefficient, noisy, and prone to vibration. The optimal design is exemplified by the ubiquituos involute spur gear:
The blue lines shown on this picture are the theoretical pitch circles of the gears; if you replaced the gears with idealized rubber rollers of these diameters, the behavior of the transmission would not change. The linear speed of any mating gears is the same when measured along the pitch circle, and the ratio of angular speeds depends on the ratio of their pitch diameters. For example, if one of the gears has a pitch circle with a diameter of 10 cm, and it's turning another gear with a pitch diameter of 4 cm, one turn of the former will rotate the latter by 10 / 4 = 2.5 turns.
Before getting any real work done, you need to make up your mind about several key parameters that will define the gears used in a particular project; except for tooth count, these values should be the same for any two gears that are supposed to mesh with each other.
This parameter depends chiefly on the transmission ratios you are hoping to achieve. The only constrain here is that with very low tooth counts, you will see a behavior known as undercutting: the two meshing gears will collide, necessitating the introduction of some additional clearance - and thus weakening the part:
This is not a big deal if the undercut is small; but when it begins to extend significantly into the area where power transfer takes place - that is, the region above the so-called base line, the gray circle shown in the earlier animation - the performance of the gear may be compromised. To avoid having to heavily compensate for this issue, unless exceptional circumstances apply, try to keep tooth counts over 8-9 or so.
Note: most of the gear generators you can find online can't deal with undercuts at all: quite simply, if you try to specify fewer than a certain number of teeth, they will reject your input or generate an incorrect result. Avoid these like a plague - and scroll down for better alternatives.
As for the maximum number of teeth - there is no real limit to how far you can go, but to keep your models simple and the cutting process rapid, it's preferable to keep the count under 100 or so.
This parameter is measured at the pitch circle. The tooth must be wide enough to allow the mating region to be machined with the tools you happen to have at your disposal; in fact, it's advisable to use tooth width at least 30% greater than the diameter of the smallest end mill in your collection.
In precision applications that do not involve extreme torques, the value will typically range from 0.3 to 1 mm - and 0.6 mm is a good starting point. You may want to increase the width when working with larger models, to keep the number of teeth sane.
When you look back at the animation shown earlier in this section, you will notice a slanted line that goes through the center. This is the path along which the gears come into contact, and the actual torque transfer takes place. The slant of this slope can be selected to suit your needs, but it needs to be reflected in the geometry of the resulting teeth:
In general, low pressure angles result in quiet-running, precise gears that are easy to machine due to generous clearances - but will be weaker, and have more pronounced undercuts. High pressure angles result in thicker tooth profiles which need to be machined with smaller tools, and can be more noisy and prone to backlash and slippage - but also survive a lot more abuse.
Sensible values for this parameter range from 12° to 25° or so - and my personal favorite for miniature assemblies is 17.5°.
The addendum is the distance by which every tooth sticks out above the pitch (contact) circle. The recommended value that
ensures continuous and smooth torque transfer is
tooth_width * 2 / π, with some wiggle room if you are pressed for space or are running into
other constraints. Going below 70% of the recommended distance should be done at your own risk.
To minimize undercuts when dealing with low pressure angles and low tooth counts, you can use a larger addendum on the smaller of the two meshing gears, and a smaller addendum on the other one. This technique, known as profile shift, is particularly useful for gears with fewer than 12 teeth, where it can significantly reduce undercuts without compromising machinability (right):
Of course, profile-shifted gears will not mesh correctly with regular ones - the mating gear always needs to have exactly the opposite shift. This is seldom a problem, but you need to keep this constraint in mind.
All right! Equipped with all this knowledge, you can compute the profile for your first set of involute gears. There are many different approaches to this task, but my favorite method is unique in that it comes without the limitations that plague simpler techniques - and if you are so inclined, it can be easily generalized to deal with a variety of unusual gear shapes and tooth profiles later on.
The process begins with calculating the circumference of the pitch circle for your gear:
tooth_width * 2 * tooth_count. From this, you can trivially calculate its diameter of the circle you
need to draw (
c / π). Once you have it on your screen, you
need to add a straight line tangent to the pitch circle; this will be used to build a hypothetical mating gear
with an infinite diameter. If we can mesh with this, we will be also able to mesh with any smaller gears
encountered in real life.
On this "pitch line" of our infinite gear, we can draw two lines perpendicular to it, precisely
tooth_width apart - and rotate them around the point where they intersect the infinite gear
by an amount corresponding to the pressure angle we want to maintain (18° in this example). To turn this
into a proper tooth, we need to add a horizontal line that is normally placed
addendum = tooth_width * 2 / π below;
in fact, draw the line about 5% farther down to create some clearance for any dirt or grease caught
in between gears. When done, trim off the excess - and voila:
The next step is to simply perform a simulation of how this tooth would engage and then disengage with the
actual gear we are trying to create. Recall the mention of rubber rollers: for any two meshing gears, the
linear velocity at the pitch circle should be the same - so if the infinite gear (aka rack) moves by
n millimeters, the other gear will rotate by
360° / cpitch * n.
The rest should be fairly obvious; this is the pattern you will see if you keep creating copies of the tooth at successive offsets along the rack, and then rotating them back by a matching angle around the center of the pitch circle of our gear:
The process can be automated easily in almost every CAD tool, although for the initial try, it makes sense to do it by hand; in any case, once you have this pattern generated, you simply need to trace a curve around it, replicate this curve using the polar array tool to create the preselected number of teeth, and offset the pitch circle by the addendum distance to trim off the unnecessary bits:
Try to do this exercise on your own. With rare exceptions where backlash is completely unacceptable, you should draw the outline about 0.05 mm apart from the trace left behind by the simulation; without this small amount of play, the teeth will engage on both sides simultaneously, producing unnecessary friction.
The beauty of the simulation-based method of constructing gears is that it's WYSIWYG, and that it takes most of the usual constraints out of the equation: for example, you can produce gears with any number of teeth, because the possibilty of undercuts is automatically accounted for. On top of that, it can be modified in several fairly intuitive ways. You can:
Use custom tooth profiles: instead of the trapezoidal tooth profile we created in our example, you can as well draw a puppy, and still get a pair of working, matching gears. It won't be particularly efficient, but it will work. I'm not good with puppies, but here's an example of a less cute freeform tooth and its generated counterpart:
More practically, you can also design gears that have an involute profile, but mesh only with a particular matching gear, by using that gear's tooth profile and its pitch circle, as opposed to a rack. This will help you reduce undercuts when other methods fail.
Create custom gear shapes: there is nothing that forces you to perform the simulation against a pitch circle - the algorithm can be easily extended to use an arbitrary "pitch curve" to generate oddball gears that produce variable transmission ratios and do other funny things.
Draw internal gears: "inverted" gears - essentially a ring with teeth pointing to the inside - are indispensable in certain types of mechanisms, including high-performance transmissions covered in section 6.4. These need to be generated to mesh with a specific gear running on the inside, using its tooth profile, and are tough to get right by other means.
Sounds more fun than it should be? You bet it does!
Tip: of course, there is no need to repeat all these steps by hand. The most convenient option is this web-based generator developed by Rainer Hessmer. The tool is very user-friendly: simply enter several parameters and get a DXF file in return. The tool properly handles undercuts, internal gears, profile shifts, and so on - so take it for a spin!
Alternatively, Rhino users may want to grab a copy of the
semi-automatic script I am using in my work. It is activated by
in the command prompt, and then answering some simple questions about the gear you want to generate. It
will then draw the appropriate circles and - if instructed to do - a standard trapezoidal tooth profile.
You can modify the profile or simply select the one produced by the script,
Hob4 again to finish the job.
The script also allows you to generate internal gears; to do that, you must provide a mating pitch circle and tooth profile extracted from a previously generated external gear. This circle needs to inside, and must be touching the top of the pitch circle for the internal gear.
Involute spur gears are remarkably versatile, easy to model, and easy to manufacture - but they are certainly not the only type of a gear you can use. Let's go over some of the more exotic but still useful options that may come handy in your work.
In comparison to planar spur gears, bevel gears have a very important property: they mesh at an angle, and therefore, can be used to change the axis of rotation. Although this goal can be achieved by other means - for example, with simplified crown gears or with worm drive, bevel gears do so with remarkable precision and efficiency.
The process of constructing a set of meshing bevel gears is considerably more complicated than what we discussed before, and involves a brush with non-Euclidean spaces; nevertheless, if you follow these instructions closely, you should be able to pull it off.
As with spur gears, the first step is to choose tooth counts, width, height, and pressure angles; you should also think about the exact gear ratio you want to achieve, as one pair of gears will be not interoperable with any other differently configured set.
Next, draw the appropriate pitch circles for both of the gears you want to create, placing them on the X-Y plane; when done, add two lines perpendicular to this plane and going through the center of each circle. Finally, rotate one of the circles (and the associated perpendicular line) in the Y-Z plane, pivoting the operation around the point where touches the other one. In our example, we will use 90° between the mating gears, but any other angle is a possibility:
With this out of the way, draw a line that starts where the two perpendicular lines meet each other,
and ends where the two circles touch each other (
Q). Next, using the Y-Z drawing plane, add a line
perpendicular to this newly created one. This final line is denoted as
To minimize clutter, we'll focus on constructing the gear profile for circle
a, and hide
the other circle for time being; the steps needed to construct a suitable gear profile for the second gear
will be analogous to what we are doing past this point, so you should have no difficulty figuring it out.
The next task is to mark two locations on the line
l that are in equal distance from point
Q; these newly created locations are denoted as
distance must be a bit greater than the addendum for your gears, but beyond that, the exact value doesn't
matter. With these two points in place, draw a circle in the X-Y plane that has the same center as circle
a, and passes through point
C; a similar circle should be drawn for
Lastly, use the loft operator after selecting the newly created circles
to build a conical surface:
The next challenge is to check the manual for your CAD application, and identify an operator that allows you
to "unroll" this surface to a construction plane; in Rhino,
UnrollSrf is what you need. Select
the surface and the original circle
a, and apply this transformation; you should get an
object resembling the one on the left:
Extract the two "broken" edges of the surface as straight curves, and draw a circle that has the same center and diameter as the arc produced by unrolling the curve we selected alongside with the cone. This will be your virtual pitch circle, on which you will have to construct a regular tooth profile using the approach outlined earlier on.
You can start by simulating the appropriate tooth profile (center, above), and then by making a planar array
to complete the gear (right). The only difference is that before making the array, you need to rotate the profile
by 90° to touch one of the extracted edges; and then make a radial array consisting of
copies that fill an angle of
360° / length_of_virtual_pitch_circle * length_of_circle_a,
rather than the usual 360°. When done, truncate the parts that stick outside the edges, and you should
be all set.
To wrap up this process, you need to position the created gear profile over the unrolled surface again, making
sure that it is aligned correctly. Next, use an operator that will map the curve back onto the original conical
surface, using the unrolled geometry as a reference. In Rhino, this is accomplished with
Just make sure that you select matching locations on the reference and destination surface, and if you did it
right, you should be able to get something of this sort:
You can delete the cone and all the helper surfaces at this point, and repeat the same procedure to generate the other gear. In the end, you should be seeing a result that resembles this:
Neat, eh? You can now create a scaled copy of each outline along the appropriate axis, using a 3D scaling operator with the origin set at the point where the axes intersect; be careful with this step, as using an incorrect origin will result in gears that would not mesh. With this step out of the way, the scaled copy and the original can be then lofted together, and each of them can be individually extruded toward any point behind the gear to create a cap surface. Few extra moves to trim and cap everything neatly, and you should have a final result in front of you:
Worm gears are an interesting animal. On one hand, they offer extremely high reduction ratios, which makes them quite attractive for robotics: every turn of the input shaft advances the output gear just by a single tooth. On the other hand, they generally boast very poor efficiency due to significant friction - 30-40% being the norm.
From the design standpoint, these gears are fairly unremarkable; the output gear is generated the same
way you would make any other spur gear, while the input gear - or more accurately, the input screw -
is essentially a rack tooth that is wrapped around a spiral path, with a
tooth_width * 2 pitch
between the turns:
Depending on the thickness of the output gear and the diameter of the input screw, the profile used to generate
the output gear may need to be corrected slightly, based on a 2D projection of the trimmed,
meshing section of the screw. You should be able to create all this with no help by now;
operators such as
in Rino are about everything you need. The only complication for CNC machining is that as opposed to
spur and bevel gears, the input screw needs a two-part mold.
Helical gears are essentialy spur gears - but (literally!) with a twist. Instead of having their teeth extruded along the axis of rotation, the top and bottom profiles are rotated in relation to each other, and then lofted to form a twisted gear. The twisting may be done in a straight line (left) or in a curved fashion (center):
The point of this modification is that each gear will now engage more gradually, thus reducing the amount of noise and vibration produced under load, this matters particularly in automotive applications and other situations where steel gears transfer high torques at high speeds. On the flip side, this geometry means that the gears act like a wedge, and exert substantial axial forces on the assembly they are coupled to; a double helical layout, also known as herringbone (above, right) works around that to a large extent.
Helical gears are relatively difficult to make using simple, CNC-machined molds; for optimal results, the mold would need to be split into four to six parts. It is easier to machine them directly using a rotary axis - but in any case, the benefits of using this class of gears are negligible in small-scale robotics work.
It is possible to adjust the speed or torque of a motor by employing just a single pair of gears, but doing so is often impractical: for example, if you have a motor that nominally runs at 10,000 RPM, and want to get down to 100 RPM - a reduction of 100:1 - you would need to mesh an input gear with about 10 teeth to an output gear with no fewer than 1,000. Assuming that the width of a single tooth is 0.6 mm, the diameter of the latter gear would be almost 50 cm - oops!
To solve this problem, it is common to build gearboxes that employ multiple stages; each stage reduces the speed by a more reasonable factor - usually between 2:1 and 5:1 - and then powers the next stage that performs the same trick again. The gains are exponential: if one stage gives us 4:1, then two will offer a 16:1 reduction; three will yield 64:1; and four would get us down to 256:1.
The most rudimentary type of a gearbox is the offset-gear arrangement shown on the left; you can check out a practical example of this design here. This approach works well, but tends to take up space; a simple modification, shown on the right, is to stack the gears alternating between two shared axles, instead (example):
Offset gear transmissions of this sort are simple and adequate for most needs; for instance, Tinybot mk III uses them exclusively. Having said that, every now and then, you may wish for something more compact and self-contained. If you are pressed for space, it makes sense to consider a planetary gearbox. There are several possible ways to design such a mechanism, but the most practical arrangement relies on a small, centrally placed sun gear coupled with the input shaft; that gear is surrounded by several planets, and these planet gears in turn roll inside a stationary internal gear known as annulus or ring:
Of course, this design makes little or no sense if the planets just run freely; the output stage is
created by creating a carriage that slides onto the planet gears, often using a sleeve bearing.
The transmission ratio of such a mechanism will be
tooth_cntsun / tooth_cntannulus
- and that's quite a good result in such a small envelope. Just as importantly, additional stages of
a planetary gearbox can be packed tightly on top of each other, resulting in a form factor that is
pretty hard to beat.
To better illustrate the practical aspects of making a planetary gearbox, you may want to check out this rendering of the design I have been using in my projects for quite a little while:
Of course, planetary gearboxes are no silver bullet; they are about 2-3 times more compact than offset gears, and outperform them in some ways, but it's possible to run into situations where an even higher ratio would be desirable - but where the poor efficiency of worm gears and other types of screw-based actuators is hard to swallow.
Of all the more exotic transmission systems, the most promising one is probably the cycloidal drive. This device easily attains ratios of 100:1 or more in a single stage by brilliantly exploiting eccentric motion of a large gear in a comparably-sized "cage":
I have a separate page describing this particular design in more detail, so I won't repeat all the information here. Rest assured, cycloidal drive systems have some substantial downsides - for example, they tend to introduce vibration, and require subminiature ball bearings to isolate the input shaft from the eccentric gear. Nevertheless, when out of mainstream options, they are worth a try.
By now, you should be well-equipped to get ahead with almost any electromechanical project you can think of; the only question is where to start. In the world of robotics, the single most important choice is probably the mode of locomotion; everything else - from the selection of sensors, to power sources, to data processing capabilities - follows from that. In the last section of this chapter, it makes sense to quickly run through the options you have.
In all likelihood, your initial projects should probably be wheeled. It is tempting to try something bolder, but until you are familiar with the dynamic properties of plastics, and have several precision robots under your belt, it's really the best choice. Some of the common approaches to wheeled robots include:
Differential steering: in this design, the two front wheels are individually hooked up to two motors; the rear wheels are typically spinning freely. When the motors are rotating at the same speed, the robot drives in a straight line; speed differences result in the chassis banking to the left or to the right. The rear wheels typically need to be equipped with rotary encoders or optointerrupters to detect and compensate for unintented fluctuations in motor speed, and to sense loss of traction.
The main disadvantage of this design is that during sharp turns, the non-powered wheels will be dragged sideways, which may lead to loss of registration with the environment. The problem can be avoided by using swivel wheels - or even eliminating the rear wheels completely and lowering the center of mass so that the device is stable with just two; but all these solutions make it much harder to maintain a robust feedback loop. That last problem may be overcome by using Mecanum wheels - but the price tag may be hard to swallow.
Car-style steering: in this arrangement, the front wheels are spinning freely; a rigid linkage operated by a motor allows them to turned left and right while a separate motor powers the rear wheels. Position sensors places on the front wheels provide a feedback loop. This is the arrangement used in Tinybot mk III.
The approach is simple and sweet, which is precisely why it's so popular in the real world. That said, not all is roses: for accurate, slip-free turns, you may still have to incorporate Ackermann linkages and a differential on the rear axle. This adds more complexity than you could be expecting.
Fully reorientable drivetrain: this approach relies on a drivetrain with a system of linkages that allows every wheel to be repositioned to accommodate motion in a particular direction. A simple and economical design of this type, using only three motors and permitting on-the-spot 360° turns, is employed in my Omnibot.
The only disadvantage of the specific approach used in Omnibot is that the robot needs to come to a stop before changing direction. In situations where this is a significant problem, additional motors may be used to independently control the position of each wheel.
Wheeled robots aside, legged creatures have some undeniable appeal. These robot designs tend to pose a range of challenges related to space management, torque requirements, obstacle sensing, and the creation of practical movement algorithms. Sure - it is relatively easy to bolt a bunch of expensive servos on top of a simple frame - but going beyond that point is usually fraught with peril. Still, if you want to give it a try, your basic choices are:
Biped locomotion: achieving a natural, stable, and useful walking gait probably requires at least 4-5 degrees of freedom per leg, packed in a very narrow envelope - as well as a bunch of gyroscopes, accelerometers, and contact or current sensors to detect problems while it's still possible to recover. Because of this, and because of the algorithmic complexity of dynamically stable biped motion, most of the designs you can see tend to cut a lot of corners; using oversized feet to achieve static stability, or executing turns in a large number of small and awkward movements, is a common pattern.
Quadruped and beyond: with four or more legs, it is relatively easy to create a design capable of moving in any direction without the need for sophisticated stability control. At least three degrees of freedom per leg are probably required to achieve efficient walking gaits (making turns, climbing obstacles, etc), but some success on flat surfaces can be achieved with only two degrees. Hexapods seem to be particularly popular with hobbyists - perhaps because of their sinister look?
Of course, a number of designs for "scuttling" robots with one or less than one degree for freedom per leg can be found on the Internet - but they are usually not any more practical than a wind-up toy.
These movement modes aside, there is a long tail of exotic inventions that are worth at least reading about. We have hopping pogo stick robots; several well-publicized ball-balancing units; a bunch of levitating designs that leverage air cushions, magnetic fields, or vibration; insanely cool crawling robots; and a large selection of flying and sailing ones. And hey - why not add something of your own to that list?
Click to proceed to chapter 7...