No-brainer: Morphological Computation Driven Adaptive Behavior in Soft Robots

Alican Mertan, Nick Cheney

TL;DR

This paper demonstrates that complex, adaptive behavior in soft robots can be achieved entirely through physical body design, without any neural network or central "brain." By evolving voxel-based robots with stimuli-responsive materials, the researchers created machines that perform "closed-loop morphological computation," effectively acting as physical logic gates and even forming swarms capable of memory-like functions.

What Problem Does This Paper Address?

In traditional robotics and AI, there is a sharp divide between the "brain" (the controller/neural network) and the "body" (the mechanical hardware). This dualism often leads to "fragile" systems where the controller is overly specialized to a specific body shape.

The authors argue that nature doesn't always work this way; many biological organisms (like plants or simple invertebrates) exhibit intelligent, adaptive behavior through their physical structure alone—a concept known as morphological computation. The challenge addressed here is whether we can evolve artificial soft robots that sense their environment and change their behavior in real-time solely through physical shape-shifting, bypassing the need for a digital or neural controller entirely.

Methods

The researchers used a 2D voxel-based soft robot simulator called EvoGym. Their approach combined several key components:

  • Reactive Materials: They designed "sensory voxels" that expand or contract in response to two binary environmental stimuli. They also included "active voxels" (muscles) that expand and contract on a fixed sinusoidal cycle to provide locomotion energy, but notably, these muscles have no feedback or "intelligence."
  • Evolutionary Algorithm: They employed MAP-Elites, a Quality-Diversity algorithm. Instead of just looking for the "fastest" robot, MAP-Elites searches for the best-performing robot for every possible behavioral niche (e.g., a robot that moves left when Stimulus A is present but right when it is absent).
  • Generative Encoding: Robot bodies were represented by Compositional Pattern-Producing Networks (CPPNs). This allowed evolution to "paint" complex arrangements of sensory, active, and passive voxels across a grid.
  • Closed-Loop Logic: The behavior is "closed-loop" because the environmental stimulus changes the robot's physical shape, which immediately alters its mechanical interaction with the ground, thereby changing its behavior (direction/speed).

Datasets

The study does not use external datasets. Instead, it utilizes the EvoGym simulation environment as a benchmark for evolving and evaluating soft robot morphologies. The "data" produced consists of an archive of evolved robot designs categorized by their behavioral responses to stimuli.

Key Insights and Ideas

  • The Body is the Computer: The robot's physical configuration is the algorithm. There is no code being executed to decide which way to turn; the physics of the expanded/contracted voxels dictates the gait.
  • Behavioral Logic Gates: The authors realized that a robot moving left or right based on binary inputs is functionally equivalent to a Boolean logic gate. For example, a robot that only moves right if both Stimulus 1 and Stimulus 2 are present is effectively a physical AND gate.
  • Complexity Costs Fitness: Evolution found it much easier to create robots that ignore stimuli (moving in one direction regardless) than robots that perform complex, non-linearly separable logic like XOR. The "smarter" the body, the slower its overall locomotion tended to be.
  • Emergent Memory from Memoryless Parts: By "wiring" robots together in a swarm—where the movement of one robot triggers the stimulus for another—the researchers hand-designed a D-type latch. This is a significant "aha moment": it shows that a collection of purely reactive, memoryless physical objects can collectively exhibit memory.

Key Results

  • Full Behavioral Coverage: The evolutionary process successfully discovered morphologies for nearly all 16 possible combinations of behaviors across two binary stimuli (e.g., Move Left/Right/Left/Right).
  • Statistical Significance: The researchers demonstrated that while "stimuli-aware" robots are slower than "dumb" ones (due to the mechanical constraints of reactive voxels), the adaptive behaviors were robust and repeatable across different bounding box sizes (5x5 to 10x10).
  • Proof of Concept Swarm: The D-type latch swarm successfully demonstrated that morphological computation can be scaled to perform any computable behavior, provided the "wiring" (environmental interaction) is correct.

Limitations and Future Work

  • Manual Swarm Design: While the individual robots were evolved, the connections between them in the swarm (the "wiring") were hand-designed. Future work could involve evolving the swarm structure itself.
  • 2D Constraints: The experiments were limited to a 2D voxel environment. Moving to 3D would significantly increase the complexity of the morphological computation but also the potential for more nuanced behaviors.
  • Binary Stimuli: The current model uses simple "on/off" environmental cues. Real-world environments are continuous, and adapting these reactive materials to gradient-based stimuli remains a challenge.
  • Hybrid Intelligence: The authors acknowledge that the goal isn't necessarily to replace the brain, but to create a "multiscale competence architecture" where the body handles low-level adaptation, leaving the brain free to handle higher-level tasks.

Summary updated

Abstract

It is prevalent in contemporary AI and robotics to separately postulate a brain modeled by neural networks and employ it to learn intelligent and adaptive behavior. While this method has worked very well for many types of tasks, it isn’t the only type of intelligence that exists in nature. In this work, we study the ways in which intelligent behavior can be created without a separate and explicit brain for robot control, but rather solely as a result of the computation occurring within the physical body of a robot. Specifically, we show that adaptive and complex behavior can be created in voxel-based virtual soft robots by using simple reactive materials that actively change the shape of the robot, and thus its behavior, under different environmental cues. We demonstrate a proof of concept for the idea of closed-loop morphological computation, and show that in our implementation, it enables behavior mimicking logic gates, enabling us to demonstrate how such behaviors may be combined to build up more complex collective behaviors.
BibTeX
@inproceedings{mertan2024no,
  author = {Alican Mertan and Nick Cheney},
  title = {{No-brainer: Morphological Computation Driven Adaptive Behavior in Soft Robots}},
  booktitle = {SAB},
  year = {2024},
  url = {https://openreview.net/forum?id=k3s3jfwnQU},
}

Loading related content...

Figure