From software intelligence to physical work: how AI is reshaping robotics

When investing we look for three things – businesses that are able to demonstrate how they can transform whatever they are focused on in a really meaningful way. The second thing is companies which have the potential for international reach. And the third thing is more about the people – backing people who are credible and will have a fighting chance of actually being able to market and sell their robots rather than just build them.

The VLA models, which are effectively the physical world equivalent of a large language model is an area we’re really active in and we’ve got a number of companies which are at the cutting edge of developing those models.

It’s an area that we have to tread quite carefully in though as I suspect we are still a number of years away from having the kind of systems which can immediately manifest themselves in great commercial results.

You have got to be quite selective in the battles you fight. Small UK companies are not going to compete in the world of humanoid robots because these things are backed by enormous tech companies or states such as China. But on the business-to-business side and industrial side, there is almost a limitless selection of applications that need to be automated and most need a bespoke system.

There are a number of niche areas where the UK is particularly strong. We’re quite active in autonomous vehicles at sea for example and we have four or five companies that are genuine world leaders. Perhaps not surprising given that we’ve got so much coast around this country.

We build the software which can intuitively be used to connect to a robot and remotely operate it from anywhere in the world. The kit has super low latency and uses virtual reality.. Anybody can use and control the robot which is a game changer for the industry. And it means robots can be deployed very quickly. Traditionally you would need very high levels of expertise but with our system you just plug and play.

We’re working with Leyland trucks in Preston where we’re deploying robots that can paint trucks and assemble parts. Jobs that are either toxic environments for humans or monotonous jobs that require very long hours.

In agriculture we have several major projects, one for strawberry picking and tomato picking and one with grape picking. These are very high value crops in an industry where it has been difficult to find workers.

We have also partnered with Nvidia because we have found that simulation provides the backbone for training robots.

In the UK there are lots of industries which are still manual, and this is because there is a very high level of variation in them, so they have traditionally avoided automation. We have smart robots to address these problems.

Our vision is to provide intelligence to robots. That transition of skills does take human involvement. With our software a skilled human can train the robot and teach them things like dextrous movements. And we can collect data from those dextrous movements and that can be millimeter level tasks such as screwing a bolt. It would be very expensive to program those. With our software because of the embodied AI machine learning technology this is achievable. On the journey of having a skilled human and then a skilled robot, there are lots of steps between these two spectrums.

Several universities have ramped up activities around robotics plus AI as a departmental focus. They are seeing robotics less about hardware and engineering and more about the software and the intelligence behind the robotics. TechUK is working with Nvidia and QA (the UK’s largest training provider) to look at offering apprenticeships to join up the pipeline so that universities have a really high calibre of students with industry experience in robotics plus AI.

It’s about applying a lot of the work that we’ve already done around digital services – standards, guardrails, assurance mechanisms – and apply them to physical systems. We are working with the likes of the Alan Turing Institute, the Open Data Institute and the Ada Lovelace Institute and we run an annual digital ethics summit.

There is very little robotics-specific regulation. But, as an example, we have a drone company that can deliver blood samples between hospitals. But for the loading onto a drone and the offloading they need to deploy a robot to trundle along the corridors to pick up the sample and take it to the clinician. To allow this, the firm comes across a whole number of different overlapping regulatory regimes, things like data collection and health and safety. The Regulatory Innovation Office (announced in October 2024) is bringing all these around the table to unpick this complex web of regulation and streamline the path to deployment for companies.

We’re pushing for more ambition. There is an immense opportunity for the UK to extend AI beyond the screen into hospitals, schools, roads, you name it. So it’s a question of how we deploy it and scale it and seize the opportunity as quickly as possible. There are huge potential gains around productivity to be had.

It’s partly about education and upskilling VCs and bringing in the institutional investors, like pension funds, the National Wealth Fund and the British Business Bank. It’s about joining up the funding pipeline too.

Real-world deployments for swarm robots need them to work out-of-the-box at scale in messy places: think homes, agriculture, construction, environmental monitoring, agile manufacturing, city logistics.

As digital AI developments have gone from AGI to more narrow and collaborative agents (Agentic AI), I hope the same will happen with the move from general purpose robots to collective of more specialised and useful robots. Perhaps we can achieve general intelligence through these narrow agents, using swarm AI.

The key benefits of swarms are that individuals are narrowly designed and so easier to implement and test. As a result these robots can be made greener and more ethically tractable than more complex or general systems.

Intelligence is embedded in each robot, giving rise to a distributed system that can scale to small or large numbers without modifying any infrastructure – just take more robots out of the box.

New tools such as VLAs (Vision Language Action) may help push swarms into the realm of real-world feasibility.

Swarm robots need on-board intelligence to operate out-of-the-box using their own perception and cognition which makes them interesting for environments that are more challenging or lack infrastructure.

Most of our swarms are simpler agents that currently don’t translate to the existing humanoid focus of VLAs. Perhaps we should be training these models for a zoo of robots types with the idea that the world will be best served by robots with different forms, rather than aiming for AGIs in humanoid form.