Facebook isn’t often thought of as a robotics company, but new work being done in the social media giant’s skunkworks AI lab is trying to prove otherwise.
The company on Monday gave a detailed look into some of the projects being undertaken by its AI researchers at its Menlo Park, California-based headquarters, many of which are aimed at making robots smarter.
Among the machines being developed are walking hexapods that resemble a spider, a robotic arm and a human-like hand complete with sensors to help it touch.
Among the machines Facebook is developing are walking hexapods (pictured) that resemble a spider, a robotic arm and a human-like hand complete with sensors to help it touch
WHAT IS FACEBOOK’S ROBOT RESEARCH LAB?
Facebook has a dedicated team of AI researchers at its headquarters in Menlo Park, California that are tasked with testing out robots.
The researchers are currently overseeing three different projects:
- A two-armed robot
- A spider-like hexapod
- A robot with tactile sensors
The hope is that their learnings can be applied to other AI software in the company and make those systems smarter.
Facebook says the experiments it conducts with physical robots can serve as helpful ‘test cases for AI’ in other scenarios, such as online environments.
Many of the robots being tested use ‘self-supervised learning approaches’ to become smarter, based on what Facebook calls a reinforcement learning algorithm.
This means they’re allowed to roam independently, teaching themselves to recognize objects or learn new abilities through a process of trial and error.
One robot that is putting this approach to work is the six-legged droid.
The robot has no prior knowledge of how to walk or what its surroundings look like, but as it learns more information, its abilities improve over time.
It tries out several different methods to start walking and whichever one gets it closer to its goal, a controller will denote it as a reward.
In turn, the robot begins to learn which methods are successful and which aren’t.
‘Learning to walk is challenging because the robot must reason about its balance, location, and orientation in space, with the help of its sensors, such as the sensors on the joints of each of its six legs (because it doesn’t have sensors on its feet),’ Facebook researchers wrote in a post.
‘Our goal is to reduce the number of interactions the robot needs to learn to walk, so it takes only hours instead of days or weeks.’
Facebook isn’t often though of as a robotics firm, but new work being done in the social media giant’s skunkworks AI lab is trying to prove otherwise. Pictured is a robotic arm being tested
Many of the robots use ‘self-supervised learning approaches’ to learn, based on what the firm calls a reinforcement learning algorithm. Pictured is Facebook’s model for the algorithm
Essentially, Facebook is trying to recreate how a human learns what works and what doesn’t in a model that can be applied for robots.
Facebook researchers are using a similar method with a two-armed robot, except instead of just using a reward-based system, they’re letting the machine be a little more curious.
‘“Curious” AI systems are rewarded for exploring and trying new things, as well as for accomplishing a specific goal,’ the firm explained.
‘Although previous similar systems typically explore their environment randomly, ours does it in a structured manner, seeking to satisfy its curiosity by learning about its surroundings and thereby reducing model uncertainty.
‘We have applied this technique successfully in both simulations and also with a real-world robotic arm,’ researchers added.
It’s not yet clear how Facebook will use these insights in the future, but it could prove to be valuable for the firm’s first AI-equipped hardware, the Facebook Portal
Other organizations are testing digital simulations to perform many of these same tasks, but they’re able to iterate much faster, since the tests are being carried out on a computer.
Doing this in the physical world is much slower, but this can eventually set robots up for much smoother deployment in real-world environments, such as on a sidewalk or in the home, where obstacles are unpredictable.
Facebook believes it might be able to use the insights gained from training its robots to help AI learn in other scenarios, like getting a digital assistant to guess what users want.
‘A lot of the interesting problems and interesting questions that are connected with AI—particularly the future of AI, how can we get to human-level AI—are currently being addressed by people who work in robotics,’ Yann LeCun, Facebook’s chief AI scientist, told Wired.
‘Because you can’t cheat with robots. You can’t have thousands of people labeling images for you.’
It’s not yet clear how Facebook will use these insights in the future, but it could prove to be valuable for the firm’s first AI-equipped hardware, the Facebook Portal.