Google Brain has a virtual robot that designs its own body parts depending on the terrain it is faced with; it is unlike anything we could imagine. In the natural world, cognitive abilities are dependent on body structure. Every animal interacts with the world based on the body it has, which in turn affects their cognition. David Ha, a computer scientist and an AI expert at Google Brain who experiments with virtual robots, says, “Evolution plays a vital role in shaping an organism’s body to adapt to its environment...The brain and its ability to learn is only one of many body components that is co-evolved together.”
Ha was curious to see if he could apply a similar process of evolution to virtual robots called walkers. He used reinforcement learning, a technique whereby the walkers were awarded points for “good” behavior, in this case being able to navigate their environment. Ha created the environment with OpenAI Gym, using virtual physics which randomly generates “natural” spaces. The walkers were tasked with improving their locomotion through the courses and were able to redesign their own legs over a couple days.
Using trial and error and motivated by the points reward system, the bots arrived at better solutions that could not be predicted. Ha says the bipedal bots being able “to learn a better version of its body obviously enables it to achieve better performance,” Most importantly, his experiment shows that embodied cognition also applies in the realm of the walkers and they were motivated to redesign themselves to better navigate their virtual world.
The virtual bots are weird and unsettling to watch, but they are learning without human input. What we are seeing are the awkward baby steps of self-learning systems that will one day create amazing solutions we could never conceive of. This is where our responsibility to steward their massive potential comes in to play.
Read more HERE.
Reality Changing Observations:
Q1. How could this ability to redesign come into play in gameplay?
Q2. What benefits would there be in a corporeal robot’s ability to redesign itself?
Q3. Why should we be goods stewards of self-learning systems?