We have long been used to the idea of the humanoid robot a robot with two legs, two arms, a torso, and a head, and with the proportions and outward appearance of a human. In the last decade we have also become used to the reality of the humanoid robot. Universities, research institutes, automobile and consumer electronics companies, and even toy manufacturers have surrounded us with demonstrations, prototypes, and commercial products, and we can see them walking on two legs, picking up small objects, and so on. Some of the more volatile minds in our society have recently begun to fret about the possibility that humankind may become dependent on, and perhaps subservient to, the descendants of this first wave, and may even be replaced by them. We dont think theres any danger of that because, with a very few exceptions such as MITs COG project, the only resemblance these robots have to real humans is external and superficial: they look like us, but their operationalprinciples are so far removed from our own that there is little prospect that their incremental refinement will result in anything better than more of the same.
What is the problem? Skin. This opaque envelope conceals its contents from our perceptual apparatus, and prevents us from paying enough attention to what lies within. We therefore overvalue human-like shape, and undervalue human-like functioning. However, when what is inside is revealed to our senses as in the brilliant plastination preparations of Gunther von Hagens we immediately and irreversibly become more aware of our true physical nature. But there is more to how we work than what we are made of, and even von Hagens process yields no more than static displays. Take what was once regarded as a key human attribute raising us above other animals bipedal walking, which all humanoid robots must of course be able to do. Do they do it like us? Many of the best modern humanoids give clues that something may be wrong. Look at the movies of them walking and ask yourself Did even Groucho Marx walk like that? And then look at a movie of Steve Collins passive dynamic walker no motors, and no electronics, and yet it clearly walks in the same way as we do! If you use the right principles, you get the right output. Our claim is that humanoid robots should be designed using human-like principles, and we believe this approach needs a new name: anthropomimetic robotics.
Here at Essex we are exploring this new field with the CRONOS series of robots, designed and constructed by Rob Knight. The inspirations for the first examples are the human skeleton and its musculature. We are all familiar with the skeleton, and to some extent with muscles, but we are much less familiar with the ways in which the muscles are connected to the skeleton and act on it to produce movements and to exert force on the external world. CRONOS, a design study, illustrates our approach. The robot has a skeleton of rigid bone-like elements (hand-moulded from the modelling polymer Polymorph) linked by joints of various kinds (cast-in ball and socket, plain, and needle-roller). Many of the bones and the degrees of freedom of the joints have been modelled using Grays Anatomy and other medical texts as sources. The muscles are servomotors, but they are connected to the bones by elastic tendons (bungee and shock cord) forming a series-elastic combination similar to the actuators used at MIT in the 1990s. (Series elastic actuators solve many of the problems caused by using conventional motors and gearboxes rigidly coupled to limbs and other robot body parts for a good summary of these issues see the Yobotics web site and the patent.) What makes CRONOS different from other robots is that the positions and lines of pull of the muscles are also similar to those of the corresponding real muscles, and the tendons are positioned in biomimetically consistent ways. For this initial investigation, not all muscles were represented some were replaced by tendons alone, appropriately tensioned.
Even before powering CRONOS up, it is clear that there is a fundamental difference between this and the conventional approach. You take his hand and shake it: it moves easily, and so does his whole skeleton. This multi-degree-of-freedom structure, supported by the tensions between dozens of elastic elements, responds as a whole, transmitting force and movement well beyond the point of contact. You take his arm and push it downward: the elbow flexes, the complex shoulder moves, and the spine bends and twists. When the robot is powered up, it will move to some equilibrium posture, but the character of the movement is again highly distinctive, because the disturbances due to the robots own movement are propagated through the structure just like the externally imposed loads. After the first encounter, your impressions will tend to coalesce in a single thought: this is a crazy way to build a robot. And so it is, if all one wants is a robot that fits into a human envelope, is able to operate in limited ways on a largely static and predictable world, and is tractable from the point of view of control. But if we want to end up with a robot that as far as possible works in the same way as a human an anthropomimetic robot we must face up to the problems that robots like CRONOS present.
Perhaps the best way to appreciate what is involved arises from a simple and well known experiment, or party trick. Make an object that looks exactly like some everyday object, but is much lighter than it should be, and ask someone to pick it up and hand it to you. We are all familiar with the effects: the persons hand shoots up into the air, because the motor program being executed was compiled in the expectation of the object being what it appeared to be. But next time you see it, dont watch the hand; instead, watch the rest of the body. If the object is the real thing, the body will scarcely move at all, but if it is the fake, there will be lots of movement. Why? Because the original motor program did not deal only with the hand and arm, but with the whole body. With the real object, as well as directing the reaching, grasping, and lifting, it would have anticipated and cancelled the effects on the rest of the body caused by the movement and the load. With the fake, the movement and load are different, and so the body makes unexpected movements, which are reactively compensated by postural and other reflexes, producing movement out of all proportion to the cause. In a rigid robot, there would be no need for any of this unless the new load and movement affected the robots balance. However, with CRONOS, we will be forced to develop complete whole-body predictive motor programs in order to produce human-like reaching, grasping, and lifting from an apparently stable platform.
CRONOS was an initial design study, investigating materials, actuators, joints, and morphology. We have now moved on to the second phase, the prototype CRONOS2. It retains all the principles of CRONOS, but uses much more powerful motors, and some engineering simplifications. For example, CRONOSs shoulder was faithfully modelled on the human, so much so that it could be dislocated in the same way; CRONOS2 has a more conventionally engineered multi-degree-of-freedom shoulder that reduces complexity without affecting the central ideas. CRONOS2 will also be the platform for studying the neck and head. The neck will be a structure similar to the spine, and the head will carry and direct the camera(s) for the visual system.
Since vision will be the robots major source of information (as in humans), we will again try to achieve a degree of fidelity in this area. The control of gaze direction is of crucial importance to any visual system; for this, most robot implementations rely on a basic pan-and-tilt camera rigidly mounted on a rigid body structure. Gaze direction control using such an arrangement reduces to a rather crude analogue of eye movements alone, and is simply a matter of issuing the appropriate pan and tilt coordinates. However, in ourselves and in the robot, gaze direction involves not just eye movements, but also head and body movements and of course the characteristics of the last two are those of the CRONOS concept, potentially involving the whole elastic body in a nexus of predictive corrections in order to achieve stability.
The picture left shows the first prototype of the neck, head, and eye system mounted on CRONOS2. In order to provide more mobility, the neck is rather longer than the human equivalent, so much so that the robot could inspect its navel if it had one. The single eye consists of a 640x480 colour CCD sensor mounted in a spherical analogue of an eyeball which is moved by analogues of human eye muscles to give pan, tilt, and rotation. The foveal acuity of the current system is at about cat level - around 20% of the human level - but the total field of view is rather narrow, at around 25 degrees. (The eventual camera system will provide increased foveal resolution and a field of view of 90 degrees.) In order to evaluate the prototype, eye movements will initially be controlled using a saliency mapping system.
We're also working on the hips and legs. The picture on the left shows the initial design study, aimed at investigating and exploiting intrinsically stable structures for standing and walking. Once again, anthropomimetic design produces engineering benefits. The knee joint, seen here without the patella, combines rolling and sliding motions, delivering increased mechanical advantage when bent, and a distinct locking characteristic when straight. The foot, with a fixed arch out to the little toe and a jointed and tensioned big toe, works in conjunction with the ankle both to adapt to the terrain and to stiffen under load. The entire assembly is passively stable when the legs are straight, and can safely be deflected at the hip for a couple of inches. The next step will be to power the relevant degrees of freedom for controlling walking; we believe that the passive stability already obtained, combined with the use of the passive dynamic principles mentioned above, will deliver low power walking with some intrinsic adaptation to the terrain.
We are still feeling our way in this territory, and we'd be happy to receive any comments on the approach, or new information about similar projects, whether humanoid or not for example, our ideas have lots in common with Peter Dilworths project on the robotic dinosaur Troody, and there is some overlap with Aaron Edsinger-Gonzales' DOMO at MIT.
There's more information about the construction of the latest version here, in a paper presented at the AISB Symposium on Biologically Inspired Robotics. The picture below shows what the robot looked like in late 2006; here is a 6MB movie of us trying out its brand new shoulder-blade-equipped left arm via remote control. More videos can be found here.
The website is maintained by Richard Newcombe.