(Nanowerk News) When manipulating the arcade claw, the player can plan all he wants. But once he pressed the joystick button, it was a game of wait and see. If the claw misses its target, he has to start over for another chance to get the bounty.
The slow, deliberate approach of arcade claws is similar to that of a sophisticated pick-and-place robot, which uses high-level planners to process visual images and plans a series of moves to retrieve an object. If the gripper misses its mark, it returns to square one, where the controller must chart a new plan.
Wanting to give the robot a more agile, human-like touch, MIT engineers have now developed a gripper that grips reflexively. Instead of starting over after an unsuccessful attempt, the team’s robot adapted in the moment to reflexively roll, squeeze, or pinch objects to get a better grip. It can make these “last centimeter” adjustments (a riff on the “last distance” delivery problem) without involving a higher level planner, much like how someone might grope in the dark for a bedside glass without much conscious thought.
This new design is the first to incorporate reflex into the robot’s planning architecture. For now, this system is a proof of concept and provides a general organizational structure for incorporating reflexes into robotic systems. Going forward, the researchers plan to program more complex reflexes to enable agile and adaptable machines that can work with and between humans in ever-changing environments.
“In the environment where people live and work, there is always uncertainty,” said Andrew SaLoutos, a graduate student in MIT’s Department of Mechanical Engineering. “A person can put something new on the table or move something in the break room or add extra dishes to the sink. We hope that robots with reflexes can adapt and work with these kinds of uncertainties.”
SaLoutos and colleagues will present a paper (“Towards Strong Autonomous Grasping with Reflexes Using High Bandwidth Sensing and Actuation”) on their design in May at the IEEE International Conference on Robotics and Automation (ICRA). His MIT co-authors include postdoc Hongmin Kim, graduate student Elijah Stanger-Jones, Menglong Guo SM ’22, and mechanical engineering professor Sangbae Kim, director of the Biomimetic Robotics Laboratory at MIT.
High and low
Many modern robotic grippers are designed for relatively slow and precise tasks, such as reassembling the same part over and over again on a factory assembly line. This system relies on visual data from onboard cameras; processing that data limits the robot’s reaction time, especially if it needs to recover from a failed grip.
“There is no way to short-circuit and say, oh shoot, I have to do something now and react quickly,” said SaLoutos. “Their only recourse is to start again. And it takes a lot of computational time.
In their new job, Kim’s team built a more reflexive and reactive platform, using fast, responsive actuators they originally developed for the cheetah mini group — agile, quadrupedal robots designed to run, jump, and adapt their gait quickly to a variety of situations. terrain type.
The team’s design includes a high-speed arm and two lightweight fingers with multiple joints. In addition to a camera mounted at the base of the arm, the team incorporated a special high-bandwidth sensor in the fingertip that instantly records the force and location of any contact as well as the finger’s proximity to surrounding objects more than 200 times per second.
The researchers designed the robot’s system in such a way that the high-level planner initially processes the visual data of a scene, marking the object’s current location where the gripper should pick up the object, and the location where the robot should place it. Then, the planner establishes an arm path for reaching and gripping the object. At this point, the reflexive controller takes over.
If the gripper failed to grip the object, rather than backing off and starting again like most grippers, the team wrote an algorithm that instructed the robot to quickly perform one of three gripping maneuvers, which they called “reflexes,” in response to a real-time measurement of the fingertip. . Three kicking reflexes within the robot’s last centimeter approach an object and allow fingers to grab, pinch or drag the object until it has a better grip.
They programmed reflexes to do without involving high-level planners. Instead, reflexes are set at a lower level of decision making, so they can respond as if on instinct, rather than having to carefully evaluate a situation to plan the optimal improvement.
“It’s like how, instead of having a CEO micromanage and plan every single thing in your company, you build a system of trust and delegate some of the tasks to lower-level divisions,” says Kim. “It may not be optimal, but it helps companies react faster. In many cases, waiting for an optimal solution makes the situation worse or irreversible.”
Cleaning through reflex
Tim demonstrated gripper reflexes by cleaning up cluttered shelves. They arranged various household items on shelves, including bowls, cups, cans, apples and a bag of coffee grounds. They demonstrated that the robot could quickly adapt its grip to each object’s particular shape and, in the case of coffee grounds, softness. From 117 attempts, the gripper quickly and successfully picked up and placed an object over 90 percent of the time, without having to back down and start over after a failed grip.
The second experiment shows how the robot can also react at that moment. When the researcher shifted the position of the cup, the gripper, while having no visual update of the new location, was able to readjust and essentially feel around until it felt the cup in its grip. Compared to a basic grip controller, the gripper reflex increases the area of successful grip by more than 55 percent.
Now, engineers are working to incorporate more complex reflexes and grip maneuvers into the system, with a view to building a general pick-and-place robot capable of adapting to messy and changing spaces.
“Retrieving a cup from a clean table—a special problem in robotics was solved 30 years ago,” notes Kim. “But more common approaches, such as picking up a toy in a toy box, or even a book from a library shelf, have yet to be solved. Now on reflex, we think we will one day be able to pick and place in every way possible, so that the robot has the potential to clean the house.”