Li Feifei’s team’s new work was published in a sub-Journal of Nature: Would AI be smarter if it had a body?
If AI has a body, will it become smarter?
The answer is yes.
Recently, a team led by Stanford University professor Li Feifei found that body shape affects the adaptation and learning ability of the virtual creature Unimal in a complex environment, and the complex environment will also promote the evolution of morphological intelligence.
The research titled "Embodied intelligence via learning and evolution" appeared in the journal Nature Communications.
"We usually focus on how AI realizes the functions of neurons in the human brain, but it would be a completely different paradigm to think of AI as something with a physical entity," Li Feifei said.
Regarding this research, Agrim Gupta, the first author of the paper, told The Paper (www.thepaper.cn), "The ultimate goal will be to have physical agents that live in a world'created by and for mankind'."
"Now, we have seen that artificial intelligence has made great progress in the research of vision and language, and will reduce the cost of doing things in the'virtual/network' world. But real things like helping the elderly or helping humans do dangerous jobs. The technology that affects human life in a good way is still out of reach. So when we have agents with physical entities, the impact will be huge and hopefully make humans better."
The research team created a computer-simulated "playground", where arthropod-like agents called "unimals" (short for universal animals, pronounced "yoo-nimals") learn and are subject to mutation and natural selection. Influence.
The simulation of each environment starts with 576 unique unimals, which consist of a "sphere" (head) and a "body" composed of a different number of cylindrical limbs arranged in various ways.
Each unimal perceives the world in the same way and starts with the same neural architecture and learning algorithm.
In other words, all unimals started their virtual lives with the same level of intelligence-only their body shapes are different.
In the learning phase, Unimal must move a block to the target location on a changeable terrain, the terrain has different difficulties-flat terrain, blocky ridges, steps or smooth hills.
The team uses a tournament-style Darwinian evolution scheme. Each unimal trained in the same environment/task combination participates in the competition with three other unimals, and the winner is selected to produce a single offspring.
The offspring undergoes a single mutation involving changes in limbs or joints before facing the same tasks as their parents.
All unimals (including the winner) have participated in multiple competitions and will only age with the appearance of new offspring.
After training 4,000 different forms, the researchers ended the simulation.
At that time, the unimals that survived in each environment experienced an average of 10 generations of evolution, and they had a variety of successful forms, including bipeds, tripods, and quadrupeds, with or without arms.
After completing this evolution 3 times in each environment (training 4000 different forms), the research team selected the top 10 best performing animals from each environment and trained them to perform eight new tasks from the beginning. For example, bypassing obstacles, manipulating a ball, or pushing a box to tilt upward.
Finally, it was found that the unimal in the changing terrain evolved better than that in the flat terrain, and the unimal that manipulated the box in the changing terrain performed best.
After 10 generations of evolution, the most successful unimals have adapted very well in terms of morphology, and they spend half of the time learning the same task as the earliest generation.
This is also consistent with the hypothesis put forward by the American psychologist James Mark Baldwin in the late 19th century. He speculated that the ability to learn things with adaptive advantages can be passed on through Darwin's natural selection.
Gupta explained, "Naturally choose the physical changes that can get the dominant behavior faster."
Since agents that evolve in more complex environments can learn new tasks faster and better, Gupta and his colleagues believe that allowing an embodied agent to evolve in increasingly complex environments will be developed in the real world. Robots performing multiple tasks provide inspiration.
Gupta said, "Humans don't necessarily know how to design robot bodies for strange tasks, such as crawling over nuclear reactors to extract waste, providing disaster relief after earthquakes, guiding nanorobots to move within the human body, including doing household chores such as washing dishes or folding clothes. Maybe the only one. The way out is to allow evolution to design these robots."
In the face of questions about how this research will specifically help the realization of these tasks, Gupta told The Paper, “Another way to look at this research is no longer to create intelligent algorithms, but you can have The correct physical form can easily complete the task at hand. In this sense, you can theoretically optimize the form of the robot that folds clothes."
For the next research plan, Gupta revealed that the current work almost only involves the surface, and the current simulation environment and learning behavior are still very simple. "We hope to expand the scope and at the same time, expand the current one agent into multiple Agents are also an interesting direction."
The Paper Journalist Shao Wen