In traditional robotics, the "brain" (the controller) and the "body" (the morphology) are treated as separate entities. The brain calculates trajectories, processes sensor data, and sends commands to a passive body. However, a growing body of research suggests that this division is inefficient. Instead, researchers are exploring morphological computation: the idea that a robot's physical shape, material properties, and mechanical dynamics can perform functional tasks that would otherwise require complex software.
The Body as a Computational Resource
The foundational shift away from "disembodied" logic began with researchers like Rodney A. Brooks, who argued in Alternative Essences of Intelligence that true intelligence emerges from being a body in a physical world. Rather than calculating complex kinematics, an agent can exploit the natural physics of its body—such as using self-adaptive oscillators to match the dynamics of a swinging arm—to achieve stable control through mechanical coupling.
This concept, often called "embodiment as computation," suggests that clever hardware design can physically solve problems that are computationally expensive for software. As Clemens Eppner and colleagues demonstrated in Lessons from the Amazon Picking Challenge, a thin, flexible suction gripper can navigate constrained environments by physically conforming to gaps. This offloads the burden of collision avoidance from the path-planning software to the mechanical design of the end-effector.
Offloading Control to Physical Mechanics
Morphological computation works by "offloading" the complexity of real-time interaction to the robot's physical substrate. This is particularly evident in bio-inspired systems:
- Frictional Intelligence: Poramate Manoonpong showed in Enhanced Locomotion Efficiency of a Bio-inspired Walking Robot that using materials with directional friction (like shark skin) allows a robot to climb steep inclines passively. The material itself "computes" the necessary grip, increasing stable slope angles from 25 to 77 degrees without any change to the control code.
- Passive Stability: In Musculoskeletal simulation of limb movement biomechanics in Drosophila melanogaster, Pembe Gizem Ozdil and Auke Ijspeert found that passive joint properties like stiffness and damping make motor control easier to learn. The body handles the "messy" dynamics of impact and vibration, allowing the nervous system to focus on high-level goals.
- Mechanical Logic: Systems like the self-folding structures described by Bo An in An End-to-End Approach to Self-Folding Origami Structures encode the folding sequence directly into the material layout. The structure "decides" how to fold based on environmental stimuli (like heat) rather than electronic signals.
Physical Reservoir Computing
A more radical application of this idea is treating the entire physical body as a "reservoir" for computation. In Information processing via physical soft body, Kohei Nakajima and Rolf Pfeifer demonstrated that the nonlinear dynamics of a silicone arm—its inertia, elasticity, and fading memory—can emulate complex systems.
By simply reading the physical state of the arm as it interacts with the environment, the robot can perform temporal tasks "for free." This approach, also utilized by Noel M. Naughton in Neural reservoir control of a soft bio-hybrid arm, treats physical compliance as a computational asset. It allows for highly efficient control (up to 75 times more energy-efficient than traditional CPUs) because the "calculation" of movement is performed by the material's own deformation.
Co-evolving Brain and Morphology
Because the body simplifies the controller's job, the two must be designed together. This is the focus of Embodied Co-Design (ECD). Xueqian Wang and colleagues argue in Embodied Co-Design for Rapidly Evolving Agents that optimizing morphology can significantly reduce the requirements for the "brain."
Researchers like Charles B. Schaff have used reinforcement learning to jointly optimize design and control. By restricting the controller to simple open-loop signals, the learning process is forced to embed stability and adaptation into the physical geometry. The result is robots that move faster and more robustly because their intelligence is not just in their code, but in their very fabric.