29th December 2023 - Jax Floro

Over 6 million years, nature procured some simple single-celled organisms into complex ones that can carry out complicated tasks. Every species has embodied smarts. The design of our physical being corresponds to the tasks that we can perform. These smarts enabled various species to survive in often challenging environments.

This raised some questions for a group of scientists from Stanford. They considered adaptability in the content of more intelligent AI’s. They asked a compelling question: how can scientists make use of embodiment to create smarter AIs?

To answer this, they set out to create an evolutionary playground to understand how the environment and morphology of an agent can affect the intelligence of AIs.

Experiment Design

In a computer-simulated evolutionary playground, they inserted nature-inspired arthropod-like agents, called “unimals” or universal animals, which undergo mutation and natural selection in different environments.

To understand the evolution of embodied intelligence, they implemented a variety of body shapes and the performance of task. Each unimal contained a sphere as the head and body of cylindrical limbs joined together in various arrangements. The researchers simulated the process of learning and the unimals learned to manipulate their bodies and survive through reinforcement learning.

Phase 1

The first phase of the experiment was navigation or walking. Each unimal started out with the same neural architecture and access to learning algorithms. Interestingly, the unimals developed numerous ways to mobilize. Some adapted a falling forward locomotion gait, some developed reptilian-like locomotion, and others learned to slither like an octopus on land.

Phase 2

The next phase of the experiment contained a more rigid form of navigation. Each unimal entered a new learning phase. It either navigated flat terrain or a more challenging terrain that included blocky ridges, stairsteps, and smooth hills. Others carried out much more complex tasks, for example, moving an object to a designated location.

Phase 3

After phase 2, the researchers created a tournament with three animals competing against each other. The winner was selected to produce a single offspring that underwent a single mutation involving changes to limbs or joints before facing the same tasks as its parents.

After training 4000 morphologies, the researchers put the simulation to a halt. Throughout the research, the surviving unimals underwent ten generations of evolution and produced a variety of morphologies.

The Gladiator Challenge

The researchers selected the top 10 performing unimals from each environment. The selected unimals performed eight new and more strenuous tasks such as obstacle navigation, ball manipulation, and box-pushing.

The results were astounding. Unimals that trained in a terrain outperformed unimals that survived on a flat surface. However, unimals that went through variable terrain (pushing the box or ball in a landscape) performed the best. The most successful unimals also learned faster individually and across generations. Some morphologies become so adapted that they grasped the task half the time their ancestors did.

The experiment showed that not only did unimals learn faster, but the various processes they underwent added body appendages, and also allowed them to adapt faster.

Building Smarter AIs

The simulation shows that agents adapt and recalibrate in more complex environments. Not only do agents learn new tasks faster, but also perform better. Complex environments provide insight for building machines that can perform more complex tasks in a practical setting.

There are many obstacles that humans are yet to overcome in building more intelligent AIs. Soon, humans might be able to design robots that will replace people performing dangerous tasks.

Surya Ganguli, Associate Professor of Applied Physics at Stanford Institute says: “To our knowledge, this is the first simulation to show that what you learn within a lifetime can be sped up just by changing your morphology”.

Key Takeaways:

• Researchers from Stanford University establish the importance of morphological intelligence and natural selection to build smarter AIs.
• They created Unimals that learn, mutate, and evolve in a simulated environment to suggest new ways to optimize embodied AIs.
• With development, these AIs will hypothetically evolve into much more complex ones to perform complex tasks with practical applications.

 

Jaclyn-Mae Floro, BCompSc

Contact W3IP Law on 1300 776 614 or 0451 951 528 for more information about any of our services or get in touch at law@w3iplaw.com.

Disclaimer. The material in this post represents general information only and should not be taken to be legal advice.