From an AGI (artificial general intelligence) point of view, the limitations of current robots are particularly severe. Robot vision is fairly advanced, but all commercially available robots currently lack a decent sense of touch. The robots’ rigid “skin” doesn’t feel anything! And the robots can’t feel the insides of their bodies either — no kinesthesia! Human eye-hand coordination combines vision, muscle movement and kinesthesia in exquisite ways; human walking relies intimately upon both kinesthesia and touch (as the recent rage for “barefoot shoes” highlights).
Nearly every part of the human body is both a sensor and an actuator, in some sense. But today’s robots are mostly inert, inactive and non-responsive material, with a few sensors and actuators tacked on here and there. The human body is a massively parallel, intelligent, self-organizing system whose intelligence gloriously synergizes with that of the human brain. Current robot bodies are, well, “robotic” in the secondary sense that word has come to hold: stiff, graceless and non-reactive. They move in a manner indicating the fact that the various parts of their bodies aren’t really aware of how they’re moving, and hence can’t adapt their movements fluidly and contextually.
We generally take for granted that robots must be this way, at least for the near future — but is this really true? Sometimes a field adopts a technology direction due to reasons of chance or short-term convenience, and then gets locked into this direction when other possibilities would have been equally feasible, and perhaps preferable in a long-term sense. Perhaps there are other, radically different paths for robotics incorporating more aspects of biological bodies that would serve some critical purposes (such as artificial general intelligence) better in the coming years and decades.
The only ways to create robot bodies truly closely analogous to human bodies would be to emulate molecular biology, or else to develop some other sort of nanotechnology with comparable complexity. These are viable and laudable research approaches, but require a lot of preliminary steps of uncertain difficulty and duration before they can yield practical results for robotics. However, there may be alternatives that lie, in some senses, halfway between molecular biology-based robots and current inflexible “lifeless” robot bodies. One class of such alternatives, which I’ve been thinking about for a while, is “inorganic macrocell bots” or Imbots.
Imbots, as I intend the term, would be robots made of multifunctional, flexible “balls”, that may be very loosely thought of as large-scale “cells” composing the robot. The macrocells could be various sizes, but what I’ve thought most about is the case of macrocells with radii between, say, a quarter centimeter and a centimeter.
Perhaps the most critical sort of macrocell would be what I think of as a “tactile / muscle ball” (“TM cell” for short). A simple form of TM cell would have the following properties:
Perhaps the most critical sort of macrocell would be what I think of as a “tactile / muscle ball” (“TM cell” for short). A simple form of TM cell would have the following properties:- Spherical by default
- Containing a programmable microprocessor
- Able to rapidly deform into varyingly elongated ellipsoidal shapes, based on its programmed instructions
- Able to connect to other balls (including TM cells), in a way that transmits or obtains electricity and/or a stream of bits with them
- Able to sense pressure on any part of its external surface, and rapidly register this sensation in its processor’s memory
By connecting together TM cells, one could potentially make a robot able to move about in a quite flexible and responsive way. Note that every part of the robot (at least, every macrocell, not the insides of the macrocells) would be able to both act and feel. So the robot would have sensors all over the outside of its body, and inside its body as well. If a robot of this kind walked, for example, it would not do so by the sole activity of a small number of joints (like current robots), but rather by the coordinated stretching of a host of interconnected balls, which could adapt their expansion and contraction based on the pressure they sensed from various directions.
TM cells could be augmented by various other sorts of specialist balls such as eye balls with cameras in them, ear balls, speaker balls, power balls containing batteries or connectors to external power sources, roller balls allowing wheeled movement, etc. Some of these other balls might be rigid in shape, but it would be preferable if they were also pressure-sensitive all around.
Certainly, the creation of TM cells and their kin would require substantial engineering effort. However, it’s not clear that this would exceed the effort required to incrementally improve conventional robotics in the direction of increased AGI-friendliness and/or biologicality. Conventional robotics is based on the engineering of increasingly high-precision components, which is certainly desirable for many industrial robotics applications. However, Imbots may be able to achieve impressive functionality with components displaying relatively imprecise behavior, due to the emergent processes arising when autonomously sensing/acting components are appropriately networked together.
Conventional robotics has often aimed toward a sort of “robotic precision” that biological organisms generally don’t have and that simpler incarnations of Imbots would likely also not possess. However, there are multiple obvious and subtle evolved synergies between the imprecision of biological bodies and the intelligent complexity of biological bodies. If Imbots can emulate this synergy at an abstract level, without emulating the particulars of biological bodies, then they may be able to achieve human body-like behaviors more readily than conventional robots, in spite of their lesser component-wise precision. It’s not hard to envision hybrid robots as well, combining the flexibility of inorganic macrocells with the precision of conventional robotics.
For instance, we humans are very poor at rotating our joints via precisely specified degrees. Try rotating your shoulder 30 degrees theta and 17 degrees phi; your elbow 130 degrees theta and 210 degrees phi and your wrist 23 degrees theta and 177 degrees phi — how accurately can you do it? By contrast, some current robots are great at this, enabling them to be programmed to carry out certain manufacturing tasks consistently and effectively.
However, humans can often achieve fairly precise movements without this sort of accuracy, via eye-body coordination and proprioception. This enables humans to carry out many manipulation tasks current robots can’t, in spite of their incredible precision at executing individual movements. A hybrid system might contain, for example, arms built for high precision according to a conventional robotics methodology, combined with hands and fingers built using macrocells (and designed to operate via proprioception and eye-hand coordination).
All this may sound quite speculative, but — as my colleague Joel Pitt pointed out in a comment on a previous version of this article — there is already concrete and exciting work in this general direction going on, under the name of Claytronics. Check out the cool videos on the Claytronics site at Carnegie Mellon. I don’t know if Claytronics is the technology via which Imbots for AGI will ultimately be achieved, but it certainly seems promising. If nothing else it’s a proof of principle that this sort of direction is real science and engineering, not just wishful thinking.
My own professional interest in Imbots comes from the AGI direction and I must admit that, as an AGI researcher, I have mixed feelings about spending my time thinking about robotics. I have no doubt that human intelligence is deeply tied up with human embodiment, and that current robots lack multiple sorts of cognitive-relevant flexibility that human bodies possess. So I’m quite confident that embodying early-stage AGI systems in more flexible robot bodies would aid considerably in their cognitive advancement (although I’m not as extreme as some, who claim that human-like embodiment is a prerequisite for achieving human-level AGI; I think there may be many different paths).
However, my general research approach is to focus on the more obviously cognitive aspects of AGI and leave robotics to the roboticists. But after increasing reflection, I’ve become concerned that robotics, driven by the requirements of the industrial robotics industry, may be going in a direction that’s badly suboptimal from an AGI perspective. The notion of Imbots is an attempt to counteract this by suggesting a variety of robot that may work better in terms of giving early-stage AGI systems what they need.
An AGI system connected to an Imbot would likely experience its body as continuous with its mind, much as we humans do. It would likely relate to its physical actions and perceptions in much the same way as it relates to its mental actions and perceptions, because the greater flexibility of the Imbot body more closely resembles the flexibility of the internal mental universe (as opposed to the rigidity and sensory opacity of typical robot bodies). It would likely develop a sense of the relationship between itself and its surroundings more similar to the human sense of self and environment, because its macrocells would provide it with sensitive whole-body tactility. Its sense of empathy and emotional connection would more closely resemble that of humans, because it would be able to achieve pleasure from touching and being touched in a variety of ways and would be able to mirror the observed actions and perceptions of others more sensitively.
Of course, exploiting an Imbot’s potential for AGI would require appropriate software for cognition, perception and action — but my feeling is that if reasonably able Imbots were developed, the appropriate software would emerge in synergy, derived from current proto-AGI software projects (such as the OpenCog project that I co-founded). Current robots have naturally developed in synergy with engineering control theory and Imbots would naturally develop in synergy with more flexible, adaptive-learning approaches to machine intelligence.
At this point Imbots are just an idea, but part of my reason for writing this article is in the hopes of spurring R&D pointing in an Imbot direction. Imbots would seem ideal for a DIY hardware approach, although efforts from large, well-funded corporate or government research labs would certainly also be valuable.