Image: MIT/Courtesy of the researchers
Using a technique known as smoothing, scientists at MIT have been able to make AI robots perform physical tasks better.
While humans find tasks that require dexterity relatively easy owing to millions of years of evolution, even the most advanced robots can find holding and manipulating an object challenging.
Each point of contact between the object and the robot needs to be reasoned within the machine’s algorithm. With billions of potential contact events, the task is exceptionally difficult for most robots to execute seamlessly.
However, researchers based at the Massachusetts Institute of Technology (MIT) have found a way to simplify this process of robot dexterity, known as contact-rich manipulation planning.
“Rather than thinking about this as a black-box system, if we can leverage the structure of these kinds of robotic systems using models, there is an opportunity to accelerate the whole procedure,” said HJ Terry Suh, an electrical engineering and computer science graduate student.
Using an AI technique called smoothing, Suh and the team were able to summarise many contact events into a smaller number of decisions. This enables even a simple algorithm to quickly identify an effective manipulation plan for the robot.
Suh is a co-lead author of the paper which was published this week in IEEE Transactions on Robotics. Other authors of the paper are Tao Pang, Lujie Yang and Russ Tedrake.
Much like humans, robots learn using a technique called reinforcement learning, which happens through trial and error with a reward for getting closer to the goal.
But because there may be billions of potential contact points that a robot must reason about when determining how to use its fingers, hands, arms and body to interact with an object, this trial and error approach requires a great deal of computation.
“Reinforcement learning may need to go through millions of years in simulation time to actually be able to learn a policy,” explained Suh, whose work was partly funded by Amazon, MIT Lincoln Laboratory, the National Science Foundation and the Ocado Group.
On the other hand, if researchers specifically design a physics-based model using their knowledge of the system and the task they want the robot to accomplish, that model incorporates structure about this world that makes it more efficient.
But Suh and the team found physics-based approaches to not be as effective as reinforcement learning when it comes to contact-rich manipulation planning. Instead, after conducting a detailed analysis, they found that smoothing enabled better performance.
“If you know a bit more about your problem, you can design more efficient algorithms,” added Pang, who is a roboticist at the Boston Dynamics AI Institute.
For instance, decisions around the movement of other robot fingers that are not in contact with the object are irrelevant to the task. Smoothing “averages away many of those unimportant, intermediate decisions” and allows robots to focus on what’s important.
“The same ideas that enable whole-body manipulation also work for planning with dexterous, human-like hands,” said Tedrake, an MIT professor and member of the Computer Science and Artificial Intelligence Laboratory.
“Previously, most researchers said that reinforcement learning was the only approach that scaled to dexterous hands, but Suh and Pang showed that by taking this key idea of (randomised) smoothing from reinforcement learning, they can make more traditional planning methods work extremely well, too.”
10 things you need to know direct to your inbox every weekday. Sign up for the Daily Brief, Silicon Republic’s digest of essential sci-tech news.
