Digit is a five-foot nine teal coloured robot with legs that look they have been bolted on backwards and arms that swap out, depending on the task that’s required.
While it seems a little clunky, trotting onto the stage at Web Summit this week like the half man half goat character Mr Tumnus from The Chronicles of Narnia, the Agility Robotics droid, which weighs in at 72kgs, has been designed to collaborate closely with human coworkers.
The droid has flat, spatula hands – designed to grip and move boxes around warehouses –while its legs look like they have been bolted on backwards – but that’s so that it doesn’t kick things (or people) over when it bends down to pick up objects.
Digit has feet rather than wheels so that it can climb stairs and go where humans go. The bot is similar in height to a human because the world’s infrastructure is designed with humans in mind.
What’s more, Digit is moving closer to the manufacturing and logics sector’s 4.0 dream: it’s fully autonomous – there’s no teleoperation involved – and the bot has been trained on a range of commercial LLMs that are used to conduct a variety of workflows, depending on the use case.
A fleet of Digit units are already working in logistics firm GXO’s warehouses, where one of its first tasks was to move tote bags around a Connecticut Spanx factory. Agility claimed that this deal represented the world’s first Robots-as-a-Service (RaaS) deployment of humanoid robots.
At Web Summit in Portugal this week Digit appeared on stage with Agility’s irrepressible CEO Peggy Johnson, who announced a second deployment of the technology at the Schaeffler Group – which wants to use “a significant number” of the humanoids in its global network of 100 plants by 2030.
Schaeffler has also become a minority investor in the company.
Johnson, who not so long ago was espousing the benefits of Magic Leap’s augmented reality headsets, claims that factories and logistics firms require humanoids to take over repetitive “back breaking” tasks like moving loads around and sorting things into different piles for hours at a time.
“It will make life easier for human workers,” she claims. “What we’re focussed on is augmenting humans,” she says. “Typically, this is part of their job, not their entire job – but it can throw out knees and it’s dirty, repetitive mind-numbing work.”
The Atlantic’s CEO Nicholas Thompson joked that he quite liked mind numbing work in his conversation with Johnson, but according to the former Microsoft exec, in the US there are currently one million unfilled vacancies in the logistics space.
“Nobody wants these roles but what this allows humans to do is free up cycles to manage,” she insists.
And with yesterday’s news that Tesla boss Elon Musk is to become the incoming Trump Administration’s efficiency adviser, these humanoid robots could be a common sight in US factories.
Musk has already announced plans to start producing and using humanoid robots in Tesla’s factories, which was perceived at the time to be part of a cost cutting drive in face of weakening demand for its cars.
Getting the proposition right is a challenge, however, as Johnson acknowledges. Hardware is, well, hard.
The company has been working on this concept for a decade and has been in customer facilities for the last three years trying to hone use cases and the robots’ design.
The good news for those working in robotics is that because processing capacity is improving rapidly, Agility can exploit a wide range of consumer LLMs to help Digit understand the world.
“We take simulation data to teach Digit new skills,” explains Johnson. “With that AI and LLM we can start to improve Digit’s semantic intelligence. So, we can give Digit commands such as new workflows.
“The robot can be doing one job in the morning and another in the afternoon, all supported by AI,” she suggests.
While Agility’s bot is AI agnostic, Johnson has observed that it executes tasks differently, depending on the LLM that it has been trained on.
Ask an Anthropic-trained Digit how to demonstrate love and it might signal that with digital heart eyes. Swap that out with a different AI model and it might line the boxes it has been moving into the shape of a heart.
Digit has one LiDAR sensor and seven cameras in its neck and waist. Some are looking downwards, checking on the bot’s position, while others are looking ahead, using perception to recognise objects, according to Johnson.
The bot also uses acoustic sensors so that commands can be given verbally. But because a factory setting tends to be noisy most of its commands are issued through an iPad.
There are some areas where humans still outperform their robotic counterparts, Johnson admits: they can go longer on a single charge, for instance.
“The charge ratio is about 4:1 – so 4 mins of work equals one minute of charge. Which is not bad but we’re moving to 10: 1 that gives facility managers much longer periods of time to put Digit to work,” she claims, adding that the bot also can plug itself in when it realises that it is running low on batteries.
And while these bots have been designed to collaborate alongside humans, for now they need to keep a safe distance because they still struggle with precise movements and complex human interactions, which can lead to unexpected errors or accidents.
However, cooperative safety in the same proximity to human co-workers is something that Agility is actively working on, with a concept solution expected to be available for demo mid next year. Johnson is optimistic that this feature will be commercialised “within the next 18-24 months.”
This all sounds promising, but how did Digit perform in practice on the main stage at Web Summit? Johnson gave the humanoid a basic laundry tasks to conduct (“something my husband could do with,” she joked).
Digit easily manged to put a grey shirt into a laundry basket. The AI model recognised the shirt, and through object recognition knew where the basket was.
However, the bot proved less effective at more complex tasks that had not previously been rehearsed. For instance, it was not able to put a green shirt on top of a grey shirt. It took about 10 seconds for Digit to process this information before picking up a pink shirt and put it in a basket.
The green shirt in question was also stripped. The bot performed better when the prompt was changed to a striped shirt, proving that prompt language still must be precise if Digit is used to conduct impromptu tasks.
A final task scooping up all three shirts and putting them in the basket also took longer than it should have, because the bot chose to pick them up one at a time, rather than scooping them up as a pile as a human would have done.
Digit also dropped one shirt and had to process for ten seconds or so before it could pick it up again and complete the task.
This demo suggests that the efficiency drive Musk is hoping to achieve in his factories may be some time off, but with increased battery power and a larger variety of swappable hands Johnson is hopeful that Agility’s robots will be able to scale.
She says that it’s less about the software or the hardware, it’s about the application.
“You can think of the hands as being a tool for whatever it is you need. There may be a way that is easiest to fold laundry, but the point is it’s swappable. We’ll aim to match the right tool with the right use case,” Johnson says.
Subscribe to our Editor's weekly newsletter
Share
