Overview
Embodied computation describes problems solved by a body's geometry, compliance, and material response rather than by a digital controller. A pair of unpowered legs walking down a slope, a wooden shutter that opens when the air dries, and a soft gripper that conforms to whatever it touches are all performing computation that would otherwise require sensors, code, and motors. The term comes from robotics, where the same idea goes by morphological computation (Pfeifer and Bongard, 2006), and it has since spread into materials science, architecture, and product design.Examples
- Passive dynamic walkers (1990), unpowered bipeds that walk down a shallow slope using leg geometry and gravity alone (McGeer)
- Octobot (2016), an entirely soft robot from Harvard's Wyss Institute whose gait is driven by a microfluidic circuit and hydrogen peroxide, with no electronics
- HygroScope (2012) and HygroSkin (2013) by Achim Menges and Steffen Reichert, wood veneer apertures that open and close with humidity and no motors or sensors
- MIT Self-Assembly Lab's 4D-printed strands (2013), which fold into a predetermined shape when dropped in water
- Digital materials at MIT's Center for Bits and Atoms, lattices of identical parts whose stiffness is set by how they are assembled
Origins
Engineered systems usually separate the body that acts from the controller that decides. Embodied computation removes that separation by designing the body so that part of the deciding happens in the interaction of material, form, and environment. The controller can then be simpler, slower, or absent.
Passive Walkers
A passive dynamic walker is a pair of legs with no motors, sensors, or control that walks stably down a gentle incline (McGeer, 1990). The gait is a property of the mechanism. Leg length, foot curvature, and mass distribution fix a stride that gravity keeps repeating. Bipeds with minimal actuation added to the same geometry used roughly a tenth of the energy per step of contemporary humanoid robots (Collins et al., 2005), and in 2011 the Cornell Ranger walked 65 kilometres on a single battery charge. Behaviour-based robots of the late 1980s had shown that competent behaviour needs no central world model (Brooks, 1991); the walkers showed that a well-shaped body needs almost no model at all.
Morphological Computation
Morphological computation is the share of a task performed by shape and material rather than by control (Pfeifer and Bongard, 2006). The idea has a formal version in physical reservoir computing: a random network of springs and masses, driven by an input, produces a set of nonlinear responses rich enough that a simple linear readout can learn from them (Hauser et al., 2011). The body does the hard part of the computation, and the controller only weights the results. Physical intelligence extends the same idea to microrobots, which cannot carry electronics at all (Sitti, 2021). The reference point is biological. A pine cone opens its scales without a nervous system, and a Venus flytrap closes in about a tenth of a second by releasing stored elastic energy, without processing a signal.
Mechanisms
A handful of physical mechanisms recur across the field. Each converts an environmental input into a shape change or a behaviour with no electronic mediation.
Compliance
A soft gripper picks up an unknown object without computing its geometry because the material deforms to fit. Pneumatic soft robots are silicone bodies with embedded air channels that bend when pressurized (Shepherd et al., 2011), and the Octobot (2016) removed the last hard component: a microfluidic oscillator fed by hydrogen peroxide replaced the controller and drove the arms for about eight minutes per fill (Wehner et al., 2016). Bistable structures hold two stable shapes and switch at a threshold, which makes them mechanical memory. A metamaterial can count the number of times it has been compressed using nothing but the sequence in which its beams buckle (Kwakernaak and van Hecke, 2023).
Hygroscopic Response
Wood swells across the grain as it absorbs moisture, so a thin veneer bonded to a stable backing curls as humidity changes. HygroScope (2012) turned this into a field of veneer apertures that breathe inside a climate-controlled case, and HygroSkin (2013) into a pavilion whose plywood apertures open in dry weather and close in damp. The Urbach Tower (2019), 14 metres tall, was built from twelve timber components that curved as they dried from flat panels and were then locked in place. Thermobimetals do the same with heat. Bloom (2011) is a sunshade of roughly 14,000 bimetal pieces that curl open as the surface warms. These are responsive environments with no wiring. One consequence is that they cannot be switched off.
Self-Assembly
4D printing adds time to 3D printing: a multi-material strand printed flat folds into a predetermined shape, the letters MIT in the first demonstration (2013), when submerged, because one material swells and the other does not. Digital materials are reversibly assembled lattices of identical parts whose stiffness and behaviour depend on arrangement rather than on the parts themselves (Cheung and Gershenfeld, 2013). Xenobots take the logic into biology: frog-cell assemblies whose shapes were designed by evolutionary search on a supercomputer and then cut from living tissue (Kriegman et al., 2020).
Applications
In architecture, embodied computation offers adaptive envelopes without the maintenance, power draw, and failure modes of motorized systems. Hygroscopic shutters, elastically bent timber gridshells, and pneumatic structures all belong here, and they share with computational morphogenesis the habit of letting material behaviour set the form. Most of the work is in robotics. Soft grippers are commercial products, and material ecology, the grading of stiffness and density within a single printed object, applies the same principle to fabrication (Oxman, 2010). The slime mould Physarum polycephalum, fed oat flakes at the positions of cities around Tokyo, grows a network that closely matches the region's rail system (Tero et al., 2010), which shows that living matter can be recruited as a substrate, a theme continued in engineered living materials. In each case, part of the work of control moves from the controller into the body.
Trade-offs
Designing behaviour into matter changes the designer's task from writing control logic to choosing material, geometry, and environmental coupling. The object responds because of what it is made of, which is one definition of material intelligence, and the result cuts energy demand and part count, since sensing and actuation are absorbed into the structure. The trade-off is fixity. Behaviour encoded in a body is set at fabrication, answers only the stimuli it was built for, and is hard to model precisely, which is why most working systems pair embodied mechanisms with minimal digital control. Cyberphysical systems engineering is beginning to formalize that division. The lineage runs to cybernetics. The homeostat (Ashby, 1948) and the tortoise robots (Walter, 1949) were built to show that adaptive behaviour could come from simple physical feedback. The digital computer later made it easier to put that feedback in software, and the older physical approaches were largely set aside.