Robots now can utilize their entire bodies thanks to AI

A process called smoothing helps robots quickly determine an efficient manipulation strategy for moving things.
Jijo Malayil
A robot attempts to rotate a bucket 180 degrees.
A robot attempts to rotate a bucket 180 degrees.


The field of robotics has been witnessing rapid advancements in recent times, with systems becoming capable of carrying out complex tasks. One thing robots still struggle with is whole-body manipulation, a skill that humans typically excel at. 

Consider carrying a big, weighty package up a set of stairs. A human being may spread their fingers and lift the box with both hands, then support it against their chest by balancing it on top of their forearms while utilizing the entire body to move the box. 

In contrast, the robot must consider every possible location on the carrier's fingers, arms, and chest where the box may touch as a contact event. Planning for this work rapidly becomes impossible due to the enormous number of possible interaction events.

As a possible remedy to this, a team of MIT researchers has developed a contact-rich manipulation planning process that streamlines this procedure. The process uses an "AI technique called smoothing, which summarizes many contact events into a smaller number of decisions, to enable even a simple algorithm to quickly identify an effective manipulation plan for the robot," said a statement by MIT. 

Fine-tuning the strategy

Robots are usually made to use reinforcement learning. This machine-learning technique encourages the robot to complete a task through trial and error with a reward for getting closer to a goal. As a result of the system's need to learn everything about the outside world by trial and error, researchers claim that this sort of learning adopts a black-box methodology.

However, this trial-and-error method necessitates a lot of processing since there may be billions of potential touch sites that a robot must consider when deciding how to utilize its fingers, hands, arms, and body to connect with an item.

“Reinforcement learning may need to go through millions of years in simulation time to be able to learn a policy," said H.J. Terry Suh, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper on this technique, in a statement. 

In comparison, the process of smoothing enables reinforcement learning to perform well. In the big picture, many of a robot's choices when choosing how to manage an item aren't significant. Few critical judgments once many irrelevant, intermediate decisions have been removed by smoothing averages.

By attempting many contact sites and then computing a weighted average of the outcomes, reinforcement learning implicitly conducts smoothing. Based on this realization, the MIT researchers created a straightforward model with comparable smoothing, allowing it to concentrate on essential robot-object interactions and forecast long-term behavior. They demonstrated that this strategy could produce complicated plans just as well as reinforcement learning. This helped cut the computation time to about a minute on a standard laptop.

Encouraging results

In simulations, they first tested their method when robotic hands were assigned tasks like moving a pen to a specified configuration, opening a door, or picking up a dish. In each case, their model-based strategy took far less time to produce the same results as reinforcement learning. When they used actual robotic arms to verify their model in hardware, they obtained comparable outcomes.

"The same ideas that enable whole-body manipulation also work for planning with dexterous, human-like hands. Previously, most researchers said that reinforcement learning was the only approach that scaled to dexterous hands, but this research showed that by taking this key idea of (randomized) smoothing from reinforcement learning, they can make more traditional planning methods work extremely well, too,” said Tedrake, senior author and Toyota Professor of EECS.

Since the model they created is based on a cruder approximation of reality, it cannot manage extremely dynamic motions, such items falling. Although efficient for more laborious manipulation tasks, their methodology is unable to provide a strategy that would allow a robot to toss a can into a garbage can, for example. The researchers are now aiming to improve their method in the future to be able to handle these extremely dynamic motions.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board