- Published
Freddo the robot walks across the office and takes a plastic bottled offered by a staff member.
asserted
Freddo → publish → member
Given that a robot recently beat Usain Bolt's 100m sprint record, it's not the most startling achievement.
asserted
it → give → record
But the speed by which Freddo has been trained to walk, recognise the bottle and grasp it is impressive.
asserted
Freddo → train → it
It took just a few minutes to develop those skills and upload them to Freddo.
asserted
It → take → Freddo
His developers say rival systems could take days to attain such skills.
uncertain
systems → say → skills
I'm at Vsim, a British start-up based in Cambridge.
asserted
I → base → Cambridge
Founders Michelle Lu and Kier Storey hope one day their software will control robots that can navigate and do useful tasks in the home and workplace.
asserted
that → hope → home
"It's a weird situation with robotics because actually the stuff that we find as humans to be incredibly difficult, like gymnastics, you can get robots to do reasonably well.
asserted
robots → find → gymnastics
The stuff that humans are really good at, like fine dexterity, is really hard in robots," Storey says.
asserted
Storey → say → robots
Freddo's skills were honed in a virtual environment, where a task can be performed in a computer simulation millions of times.
asserted
task → hone → times
Once the optimum solution (known as a policy) is found, it can be uploaded and used by the hardware - in this case Freddo.
asserted
it → know → case
Such virtual simulations are a common way to train robots.
asserted
simulations → train → robots
Tech giant Nvidia has a system called Isaac Sim which works that way - Lu and Storey both worked on an early version of it.
asserted
Lu → have → it
In 2022 they decided to set up Vsim, to build the their own training system environment and other tools.
asserted
they → decide → environment
As they were starting from scratch Lu and Storey could optimise the software to exploit the powerful computer chips used in AI, known as graphics processing units or GPUs.
uncertain
Lu → start → units
"The underlying algorithms that we were using for most of these robotic simulations they hark back to the 1970s and 1980s, but those algorithms are not really brilliant fits for GPUs," Storey says.
asserted
Storey → underlie → GPUs
Within months they realised their system could work much faster than anything they had seen before.
uncertain
they → realise → anything
"Eighteen months in and we actually have a completely functional, super high-performance simulator," says Lu.
asserted
Lu → have → simulator
The software is so efficient that it can run on the hardware carried by Freddo.
asserted
it → run → Freddo
That means the robot can run tens of thousand of simulations while it is moving around.
asserted
it → mean → simulations
"It can look about a second, or so, ahead into the future for 20,000 different kind of combinations of things that might happen," Storey explains.
And that would be vital for a robot moving around an unstructured environment like the average home.
uncertain
that → look → home
"Things outside of the robot's control, like humans, animals or even other robots, could do things that require a change of strategy.
uncertain
that → do → strategy
These unexpected events could happen very quickly and the robot needs to be able to quickly adapt to ensure its actions remain safe and on-mission," Lu says.
uncertain
Lu → happen → mission
Vsim is a start-up with 10 engineers working on its tech.
asserted
Vsim → work → tech
It dominates the market for computer chips used for AI and has a leading robotics software division, with hundreds of engineers.
asserted
It → dominate → engineers
It does not build robots, instead it has a suite of software designed to let organisations train and control robots.
asserted
organisations → build → robots
That includes virtual simulation training systems and a so-called world model, external, called Cosmos, which gives a robot an understanding of the physics of the real world and how its environment might change as it moves around.
uncertain
it → include → world
But even with the powerful computer resources available to Nvidia, the software only gives a rudimentary understanding of the real world.
asserted
software → give → world
"Manipulation, - where I just grab a bottle, that's not too hard.
asserted
that → grab → bottle
The problem is when you start doing long-horizon tasks, where I say: 'I want you to take the bottle and I want you to fill it up and I want you to go pour'," says Spencer Huang, director of product for robotics at Nvidia.
asserted
Huang → start → Nvidia
But he's confident that good progress is being made.
asserted
progress → make → ?
This year Nvidia has started using AI agents to help build virtual environments to train robots and validate whether the solutions from training work or not.
asserted
solutions → start → training
"When we talk about creating the [virtual] world and actually scanning it in - a lot of that is actually manual labour.
asserted
lot → talk → that
"We're just throwing agents at it... it's basically given us a huge workforce," Huang says.
asserted
Huang → throw → workforce
Simulation is not the only method for training robots.
asserted
Simulation → train → robots
They can also be trained by watching human or video demonstrations.
asserted
They → train → demonstrations
Rika Antonova has been working in the field of robotics since 2015 and is currently an associate professor at the Department of Computer Science and Technology at the University of Cambridge.
asserted
Antonova → work → Cambridge
Her research is focused on, external developing software and hardware that can aid robots to learn complex behaviour.
asserted
that → focus → behaviour
Antonova works with a training system called MuJoCo, owned by Google's DeepMind since 2021.
asserted
Antonova → work → 2021
It's open-source software, which means researchers can use it for free, and are allowed to tinker with the code.
asserted
researchers → mean → code
…and 9 more, not listed.
The Aporia
