AI Creatures: Evolution of Mind and Body

The Interplay of Body and Mind in Artificial Intelligence

The conventional perception of artificial intelligence often portrays it as a disembodied entity – a purely intellectual construct existing within the digital realm. However, human cognition is fundamentally linked to our physical bodies. Recent research involving virtual creatures and simulated environments suggests that AI systems could significantly benefit from a similar “mind-body” architecture.

Exploring Evolutionary Roots

Scientists at Stanford University investigated the intricate relationship between physical capabilities and mental processes, drawing parallels to human evolution. The question posed was whether the brain’s development is influenced by the body’s abilities, and vice versa. This concept, initially proposed over a century ago, is intuitively understood – a hand capable of grasping facilitates object manipulation more effectively than a less specialized appendage.

The applicability of this principle to AI development, which typically follows a more structured path, remains uncertain. Nevertheless, the core question is compelling: could an AI’s capacity for learning and adaptation be enhanced if it evolved with an integrated body from the outset?

Simulating Evolution: The “Unimals” Experiment

The research team designed an experiment mirroring the simulated environments used for decades to test evolutionary algorithms. These environments involve introducing simple, geometrically shaped creatures into a virtual space, allowing them to move randomly. The most successful movers – those covering the greatest distance – are selected, and their designs are iterated upon repeatedly.

This process eventually yields polygons capable of traversing the virtual surface. However, the researchers aimed for a more robust and variable simulation. Their goal wasn’t merely to create creatures that could walk, but to understand how they learn and whether some learn more efficiently than others.

To achieve this, they created simulations populated by “unimals” – representing “universal animals.” These creatures possessed a spherical “head” and jointed, branch-like limbs, resulting in diverse locomotion styles. Some stumbled forward, others exhibited lizard-like articulation, and still others employed a flailing, octopus-like movement.

Look at them go!. Image Credits: Stanford

Introducing Environmental Complexity

The experiment extended beyond simple locomotion. Certain unimals were “raised” in environments featuring undulating hills and low barriers, while others inhabited flat terrains. Subsequently, unimals from these differing backgrounds competed in more complex tasks, testing the hypothesis that adversity fosters adaptability.

“Almost all the prior work in this field has evolved agents on a simple flat terrain,” explained co-author Agrim Gupta to TechCrunch. “Moreover, there is no learning in the sense that the controller and/or behavior of the agent is not learnt via direct sensorimotor interactions with the environment.” He further clarified that previous iterations focused on survival rather than learning through interaction.

Results: Terrain Shapes Learning

The team then presented the unimals with tasks such as navigating new obstacles, moving a ball to a target, pushing a box uphill, and patrolling between designated points. The results were striking: unimals that had learned to walk on varied terrain demonstrated faster learning and superior performance compared to their flatland counterparts.

Image Credits: Stanford

“In essence, we find that evolution rapidly selects morphologies that learn faster, thereby enabling behaviors learned late in the lifetime of early ancestors to be expressed early in the lifetime of their descendants,” the authors stated in their paper, published in Nature.

Evolution Selects for Adaptability

The evolutionary process didn’t just favor faster learners; it actively selected body types that facilitated quicker adaptation and more efficient application of learned lessons. While an octopus-like flop might suffice on flat ground, hills and ridges favored configurations that were fast, stable, and adaptable. These versatile bodies enabled the unimals to effectively utilize their cognitive abilities, surpassing their less adaptable competitors.

Implications for AI Development

Beyond the entertaining visuals of virtual creatures, this experiment offers significant insights. The research “opens the door to performing large-scale in silico experiments to yield scientific insights into how learning and evolution cooperatively create sophisticated relationships between environmental complexity, morphological intelligence, and the learnability of control tasks.”

Consider the challenge of automating complex tasks, such as enabling a four-legged robot to climb stairs. While manual programming or a combination of custom and AI-generated movements are options, the experiment suggests that allowing the agent’s body and mind to evolve in tandem could yield optimal results.

Open-Source Availability

The research group has made all code and data freely available on GitHub, allowing others to replicate and expand upon their work. However, substantial computational resources are required: “The default parameters assume that you are running the code on 16 machines. Please ensure that each machine has a minimum of 72 CPUs.”