Fresh juice

2024-05-10

DeepMind experiments with incredibly resilient robot hand

In the cutting-edge field of artificial intelligence and robotics, pushing the boundaries often requires breaking a few pieces of equipment. This harsh reality is what drove engineers at the UK's Shadow Robot Company to develop an incredibly durable robotic hand that can withstand remarkable levels of abuse.

 

 

The new robotic hand, already being put through its paces in experiments at Google's DeepMind AI research subsidiary, represents a major leap forward in both precision and resilience for robotic appendages. Its fingers can transition from fully open to closed in just half a second and apply up to 10 newtons of force in a fingertip pinch. But what truly sets this hand apart is its ability to keep functioning even after being pummeled by hammers, smashed by pistons, and subjected to all manner of impact damage.

"We knew this hand needed to be able to survive the inevitable accidents and collisions that occur during reinforcement learning experiments with AI robots," explained Rich Walker, Director at Shadow Robot Company. "Any interaction with the real world environment carries a major risk of impact damage when you have a powerful robot flailing around trying to master new skills through trial-and-error."

Reinforcement learning, the technique Walker referenced, is a cutting-edge AI training method that allows robots to gradually learn complex physical skills through repeated practice, adjusting their movements through many thousands or millions of iterations based on feedback about what works and what doesn't. It's an incredibly powerful approach, but one that often sees robots essentially "betting against themselves" as they try out bizarre, unnatural motions in an effort to explore all possibilities.

This type of haphazard, brute-force learning process was simply too much for traditional robotic hands to handle, which would quickly become mangled and inoperable after sustaining just a few accidental impacts. The new hyper-durable hand from Shadow Robot Company finally provides a solution that can withstand this abusive training regimen long enough for the AI system to eventually learn how to move precisely and with care.

"We're essentially letting these AI robots flail around like uncoordinated toddlers as they figure out how to apply the right forces and motions to accomplish manual tasks," said Ram Ramamoorthy, a leading reinforcement learning researcher at the University of Edinburgh. "You can imagine why that would quickly destroy a normal robot hand not designed to absorb intense batterings. This new hand is a critical tool that will allow us to keep pushing the limits of what robots can learn to do through pure practice."

The hand's unique durability comes from a combination of specialized materials, precisely engineered joints, and a deliberate focus on impact resistance throughout the design process. While previous robot hands prioritized light weight, precise movements, and human-like dexterity, compromises had to be made to achieve this unprecedented level of resilience. The result is a robot appendage that can repeatedly endure forces that would disfigure a human hand.

At the DeepMind robotics lab, the new hands are already hard at work engaged in all manner of tasks - grasping and lifting objects, manipulating tools, operating switches and more. With each mistaken drop, crush or impact, the AI brain learns a bit more about applying the right forces, saving that knowledge, and nudging closer to mastery through countless trial-and-error cycles. It may sound like a brutal and haphazard process, but it is the cutting edge of modern robotics - an approach that researchers believe will ultimately allow human-level dexterity without the need for meticulous hand coding of skills.

As AI systems grow more sophisticated and robotic hardware keeps pace, scientists envision future robots handling an incredibly wide range of manual skills - setting a table, repairing equipment, even performing surgery - all through the same general reinforcement learning process. The new virtually indestructible hand from Shadow Robot Company provides a crucial building block, giving AI the opportunity to learn through unbridled real-world practice without the threat of constant repairs interrupting the training process.

"Building true robotic intelligence means letting the machines make a lot of mistakes at first," Walker stated. "We're giving them hands that can take a beating so they don't have to hold back during that difficult learning stage. Once they master using the hands through experience, they'll be able to manipulate objects in the real world with human-like dexterity."

Share with friends:

Write and read comments can only authorized users