Imagine you should carry a big, heavy box up a flight of stairs. You would possibly spread your fingers out and lift that box with each hands, then hold it on top of your forearms and balance it against your chest, using your whole body to govern the box. 

Humans are generally good at whole-body manipulation, but robots struggle with such tasks. To the robot, each spot where the box could touch any point on the carrier’s fingers, arms, and torso represents a contact event that it must reason about. With billions of potential contact events, planning for this task quickly becomes intractable.

Now MIT researchers found a approach to simplify this process, generally known as contact-rich manipulation planning. They use an AI technique called smoothing, which summarizes many contact events right into a smaller number of choices, to enable even an easy algorithm to quickly discover an efficient manipulation plan for the robot.

While still in its early days, this method could potentially enable factories to make use of smaller, mobile robots that may manipulate objects with their entire arms or bodies, reasonably than large robotic arms that may only grasp using fingertips. This may occasionally help reduce energy consumption and drive down costs. As well as, this method could possibly be useful in robots sent on exploration missions to Mars or other solar system bodies, since they might adapt to the environment quickly using only an onboard computer.      

“Fairly than fascinated with this as a black-box system, if we will leverage the structure of those sorts of robotic systems using models, there may be a possibility to speed up the entire procedure of attempting to make these decisions and give you contact-rich plans,” says H.J. Terry Suh, an electrical engineering and computer science (EECS) graduate student and co-lead writer of a paper on this method.

Joining Suh on the paper are co-lead writer Tao Pang PhD ’23, a roboticist at Boston Dynamics AI Institute; Lujie Yang, an EECS graduate student; and senior writer Russ Tedrake, the Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering, and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL). The research appears this week in

Learning about learning

Reinforcement learning is a machine-learning technique where an agent, like a robot, learns to finish a task through trial and error with a reward for getting closer to a goal. Researchers say the sort of learning takes a black-box approach since the system must learn every part in regards to the world through trial and error.

It has been used effectively for contact-rich manipulation planning, where the robot seeks to learn the perfect approach to move an object in a specified manner.

But because there could also be billions of potential contact points that a robot must reason about when determining the way to use its fingers, hands, arms, and body to interact with an object, this trial-and-error approach requires a terrific deal of computation.

“Reinforcement learning may have to undergo tens of millions of years in simulation time to truly have the ability to learn a policy,” Suh adds.

Then again, if researchers specifically design a physics-based model using their knowledge of the system and the duty they need the robot to perform, that model incorporates structure about this world that makes it more efficient.

Yet physics-based approaches aren’t as effective as reinforcement learning relating to contact-rich manipulation planning — Suh and Pang wondered why.

They conducted an in depth evaluation and located that a way generally known as smoothing enables reinforcement learning to perform so well.

A lot of the selections a robot could make when determining the way to manipulate an object aren’t essential within the grand scheme of things. As an illustration, each infinitesimal adjustment of 1 finger, whether or not it results involved with the thing, doesn’t matter very much.  Smoothing averages away a lot of those unimportant, intermediate decisions, leaving a number of essential ones.

Reinforcement learning performs smoothing implicitly by trying many contact points after which computing a weighted average of the outcomes. Drawing on this insight, the MIT researchers designed an easy model that performs the same variety of smoothing, enabling it to deal with core robot-object interactions and predict long-term behavior. They showed that this approach could possibly be just as effective as reinforcement learning at generating complex plans.

“For those who know a bit more about your problem, you’ll be able to design more efficient algorithms,” Pang says.

A winning combination

Regardless that smoothing greatly simplifies the selections, looking through the remaining decisions can still be a difficult problem. So, the researchers combined their model with an algorithm that may rapidly and efficiently search through all possible decisions the robot could make.

With this mixture, the computation time was cut right down to a couple of minute on a typical laptop.

They first tested their approach in simulations where robotic hands got tasks like moving a pen to a desired configuration, opening a door, or picking up a plate. In each instance, their model-based approach achieved the identical performance as reinforcement learning, but in a fraction of the time. They saw similar results after they tested their model in hardware on real robotic arms.

“The identical ideas that enable whole-body manipulation also work for planning with dexterous, human-like hands. Previously, most researchers said that reinforcement learning was the one approach that scaled to dexterous hands, but Terry and Tao showed that by taking this key idea of (randomized) smoothing from reinforcement learning, they’ll make more traditional planning methods work extremely well, too,” Tedrake says.

Nevertheless, the model they developed relies on a less complicated approximation of the true world, so it cannot handle very dynamic motions, comparable to objects falling. While effective for slower manipulation tasks, their approach cannot create a plan that will enable a robot to toss a can right into a trash bin, as an illustration. In the longer term, the researchers plan to boost their technique so it could tackle these highly dynamic motions.

“For those who study your models rigorously and really understand the issue you are attempting to unravel, there are definitely some gains you’ll be able to achieve. There are advantages to doing things which can be beyond the black box,” Suh says.

This work is funded, partially, by Amazon, MIT Lincoln Laboratory, the National Science Foundation, and the Ocado Group.