In traditional robotics and artificial intelligence, the division of labor is clear: the software (the "brain") does the computing, while the hardware (the "body") passively executes commands. Under this paradigm, achieving complex behavior requires high-frequency digital control loops, dense sensor arrays, and intensive computational overhead to calculate inverse kinematics, handle environmental perturbations, and plan trajectories.
Morphological computation flips this paradigm. Instead of treating a robot's physical structure as a passive shell that must be controlled despite its physics, morphological computation treats the body as an active computational resource. By embedding intelligence directly into the geometry, material properties, and mechanical dynamics of the physical system, researchers can offload complex control and processing tasks from digital software to the physical substrate itself.
Offloading Control to Material Geometry
Why compute digitally what physics can solve for free? When a robot interacts with its environment, the physical constraints of its body act as a natural filter or a zero-latency control policy.
Consider the challenge of robotic manipulation. Standard control policies rely on high-precision sensing and complex grasp planning to manipulate objects with rigid, general-purpose grippers. However, George Jiayuan Gao et al. demonstrate in VLMgineer: Vision Language Models as Robotic Toolsmiths that optimizing the physical geometry of a tool allows complex manipulation tasks to be completed using significantly simpler, more robust motion trajectories. The structural innovation of the tool effectively absorbs the task's complexity, reducing the burden on the digital motion planner.
This concept of "programming" mechanics extends deep into material design. In An End-to-End Approach to Self-Folding Origami Structures, Bo An et al. show that folding sequences and final 3D geometries can be encoded directly into the spatial layout of multi-material shrinking layers. When exposed to a uniform environmental stimulus (like heat), the material executes sequential, multi-step transformations without any onboard electronic controllers.
Similarly, adaptation to complex environments can be achieved passively. Poramate Manoonpong et al. explored this in Enhanced Locomotion Efficiency of a Bio-inspired Walking Robot using Contact Surfaces with Frictional Anisotropy. By equipping a hexapod robot's feet with a passive material exhibiting directional friction (shark skin), the robot could passively navigate steep inclines up to 77 degrees. The material properties compute the necessary holding forces instantaneously, bypassing the need for active force-feedback loops or complex software adjustments.
Traditional Control:
[Sensors] -> [High-Frequency Digital Processing] -> [Actuator Torques] -> [Environment]
Morphological Control:
[Environment] -> [Passive Material Dynamics / Geometry] -> [Adaptive Behavior]
Physical Reservoir Computing
The most literal realization of computing with matter is Physical Reservoir Computing (PRC). In digital machine learning, a standard Reservoir Computer passes a time-series input into a fixed, highly non-linear recurrent neural network (the reservoir) and trains a simple linear readout layer to map the reservoir's high-dimensional states to a desired output. PRC replaces the simulated recurrent network with a physical object.
Input Signal ----> [ Physical Substrate (Soft Body / Magnet / Cable) ]
| (Inherent Non-linearity & Fading Memory)
v
High-Dimensional Physical States
|
v
[ Trained Linear Readout ] ----> Output Prediction
This works because many physical systems naturally possess the two core requirements of a computational reservoir:
- Non-linearity: Small changes in input yield complex, high-dimensional transformations.
- Fading Memory: The current physical state of the material is a reflection of both its current input and past perturbations (due to inertia, elasticity, or magnetic hysteresis).
In Information processing via physical soft body, Kohei Nakajima et al. demonstrated that a physical silicone arm can serve as a high-dimensional computational reservoir. By exploiting the material's natural inertia and damping, complex temporal tasks can be solved by simply applying a linear readout to the physical state of the soft body.
This paradigm allows robots to perform complex internal calculations, such as state estimation, using their own structural imperfections. For instance, Masashi Hamaya et al. in Pose Estimation of a Cable-Driven Serpentine Manipulator Utilizing Intrinsic Dynamics via Physical Reservoir Computing utilized the mechanical non-linearities, cable slack, and structural deformations of a flexible manipulator as a computational resource. Instead of treating these dynamics as noise to be filtered out, the authors mapped these high-dimensional physical interactions to achieve millimeter-level pose estimation, drastically reducing digital computational overhead.
The substrate itself does not even need to be macroscopic. Oscar Lee et al. developed a task-adaptive physical reservoir using the magnetic phases of chiral magnets in Task-adaptive physical reservoir computing. By tuning temperature and magnetic fields, the physical substrate reconfigures its own non-linear properties to switch between memory-heavy forecasting or nonlinearity-heavy data transformations.
Co-Optimizing Brain and Body
Because the physical structure acts as an extension of the controller, designing a morphologically intelligent system requires a tight coupling between hardware design and software optimization. If the body changes, the computational mapping changes.
To address this, researchers leverage differentiable simulators and evolutionary algorithms to co-design morphology and control. In Neuromechanical Autoencoders: Learning to Couple Elastic and Neural Network Nonlinearity, Deniz Oktay et al. introduced a framework that uses gradient descent to jointly optimize a neural network and the geometry of a complex elastic material. By coupling the non-linearities of the neural network with the mechanical non-linearities of the material (such as buckling and stretching), they proved that materials can be mathematically "programmed" to execute specific mechanical logic functions.
When control systems are intentionally restricted, the optimization process is forced to offload intelligence to the body. In Soft Robots Learn to Crawl: Jointly Optimizing Design and Control with Sim-to-Real Transfer, Charles B. Schaff optimized soft crawling robots using purely open-loop control signals without sensory feedback. Because the software controller was incapable of reacting to the environment, the reinforcement learning pipeline was forced to embed environmental adaptation and stability directly into the physical geometry and material compliance of the soft body.
Ultimately, morphological computation demonstrates that physical mechanics should not be viewed as a constraint to overcome, but as a rich computational medium. By aligning the natural resonance of materials with the goals of a robotic system, the boundary between physical structure and digital computation dissolves.