Researchers at the Massachusetts Institute of Technology (MIT) have significantly improved robotic manipulation capabilities. Using a new AI technique called “smoothing,” robots can now carry out tasks using their entire bodies, rather than merely their fingertips. This innovation can potentially revolutionize how robots are used in factories, space exploration, and other fields.
The challenge of contact-rich manipulation planning
Manipulating objects with the whole body presents a monumental challenge for robots. They must account for billions of potential contact points on an object for every touch with their fingers, hands, arms, and torso. This process, known as “contact-rich manipulation planning,” is computationally expensive and has been a bottleneck for practical applications. Humans intuitively manage these tasks, but they become overwhelmingly complex for robots.
The solution
To simplify the problem, MIT researchers developed a new AI technique based on “smoothing,” which condenses the numerous contact events into a smaller, more manageable set of decisions. This innovation allows even basic algorithms to devise an effective manipulation plan for a robot efficiently.
“Smoothing averages away many of those unimportant, intermediate decisions, leaving a few important ones,” said H.J. Terry Suh, an MIT graduate student and co-lead author of the paper published in IEEE Transactions on Robotics.
Reinforcement learning vs smoothing
While reinforcement learning has been effective in helping robots perform complex tasks, it requires immense computational power and time. According to Suh, it learns through a “black-box” system of trial and error, often taking “millions of years in simulation time” to be effective.
Smoothing, however, offers an alternative. By carefully understanding the model and the problem, researchers were able to bring efficiency to the process. Smoothing allows the robot to focus on core interactions with objects, enabling quicker and more effective task planning.
Achieving efficiency and combined approach
Despite the advancements through smoothing, searching through even the reduced number of decisions remained a challenge. The researchers then combined the smoothing model with a searching algorithm, reducing the computation time to about a minute on a standard laptop.
The team tested their approach both in simulations and on real robotic arms, achieving performance comparable to reinforcement learning but in a fraction of the time.
Applications and prospects
The implications of this research are immense. In industrial settings, factories could replace large robotic arms with smaller, more mobile robots that use their whole bodies for manipulation tasks, thus reducing energy consumption and cost.
Additionally, the technique could prove invaluable for exploration robots sent to Mars or other celestial bodies, where quick adaptation to new environments is crucial.
However, the researchers acknowledge limitations in handling dynamic tasks, such as a robot’s ability to toss an object into a bin. The team plans to refine further their approach to tackle such challenges.
Suh emphasized that 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.
Amazon, MIT Lincoln Laboratory, the National Science Foundation, and Ocado Group partially funded the work. As MIT’s robotic manipulation techniques continue to evolve, they open up new possibilities for robotics in various sectors, proving that a deep understanding of the problem can lead to innovative solutions.
With advancements like these, robots that manipulate objects as intuitively as humans may not be far from reality.
Source: https://www.cryptopolitan.com/robots-to-use-bodies-for-object-manipulation/