Robots (not) coming over here and taking our jobs! How one UK producer reckons that humanoid robots will be tele-operated by remote migrant workers

Farms worldwide face severe challenges: casual workers are hard to find in the harvest season. Meanwhile, the human population is increasing, but productivity and yield are failing to match that demand. This is why robotic automation and AI are important to agriculture, but developing robots to pick individual crops is an enormous problem. For example, the processes needed to pick strawberries, asparagus, and apples are all completely different, which means developing bespoke solutions.

This is why, in the medium-to-long term, some technologists believe that general-purpose humanoids could offer a potential solution to farming’s automation challenges. If a robot were intelligent, dextrous, and adaptable enough to take on a variety of different tasks, then it might obviate the need to develop solutions for every crop and function, as is currently the case.

But there is another problem. While it is relatively easy to train robots in simple, controlled, and predictable environments – such as warehouses designed for automated labor – it is much harder to train them in messy, unpredictable places where there may be countless local differences.

Farms are among the most complex and changeable of these: a vast arable farm in Wyoming is a very different environment to a smallholding in rainy Dorset, or a dairy farm in the humid hinterlands of southern China. So, how to train and operate humanoid robots in such a context?

Tele-operating

Being able to tele-operate a robot from a remote location – even from overseas – could be a boon for farmers. That’s the claim by Reading, UK-based start-up Extend Robotics. Azmat Hossain is the firm’s Business Development Director. Speaking in December at a Westminster food technology policy conference, he explains:

We are bringing to the industry a completely new way of training robots and managing them, and this is specifically useful for agriculture. What our software does is enable people to log into a robot from far away – it could be from anywhere in the world, with off-the-shelf hardware and a standard internet connection – but with full virtual reality they can control the robot with ease.

On the face of it, this seems to be a simple case of tele-operation of a kind that is available for many commercial humanoids today, some of which can be trained for autonomous tasks by learning movements from a remote operator. So, what is the unique element here? And why is it relevant to agriculture?

Hossain explains that three-dimensional telepresence is valuable in agriculture precisely because of the complex variables that make farming automation so difficult to achieve with a ‘one size fits all’ solution. He says:

The robot could be any robot, and we create a digital twin of it, which can then be controlled by human interaction. What this means is we are helping farms to adopt robots at very feasible costs.

And now we have full, immersive, VR-based visualisation, which enables the operator to see fruits, plants, or the farm environment in real time, in full virtual reality, and be able to do things [via the robot] with high precision and dexterity.

But here is the key, he explains:

The difference is the operator could be sitting in the comfort of their own home, or in another remote location where there could be better, easier access than actually going to the farm itself.

So, Hossain appears to be saying that his technology means that casual workers could tele-operate robots without having to travel from overseas to work on farms directly. Is that correct? He confirms that is the case: 

With all the issues of immigration and all these different issues we are facing now, it could mean that we don't need to bring people to the UK anymore. They could operate robots from other areas.

Plus, tele-operation means robots could be kept working all the time by workers in different timezones, he says – issues with battery life aside, of course.

It would thus be a form of arms-length outsourcing, in fact, in which robots become the remote avatars of low-paid offshore workers. Such an idea makes logical, technical, and practical sense, perhaps, yet it also feels exploitative, as workers would be enabling food security and increased profits in wealthy nations, while being used as minimum-wage labour themselves.

Thinking

But might that be better than doing hard labor in the field directly, and without them having to cross continents for casual work that offers minimum job security and safety? Perhaps. It is certainly an idea that demands serious debate in policy, ethical, and human rights terms, given the uncomfortable echoes of slaves working long hours on plantations.

Indeed, it would – quite literally – de-humanise casual workers, whose presence on landowners’ property would be hidden within faceless machines. It would also seem to present significant security and privacy risks: what if a disgruntled remote worker makes their robot avatar run amok, or uses it to spy on their employer?

Those issues aside, multi-purpose, generally intelligent humanoids could be trained in this way to perform complex agricultural tasks by their remote operators, continues Hossain:

Using this interaction, we can train the robots using our software. You need enormous amount of data to train these robots, and through our system, we can enable that. So, we are not only doing the teleoperation, but also training the robots for complex use cases on farms.

Again, this makes logical sense, though it suggests that seasonal workers’ primary task would no longer be picking crops but providing high-value training and data-gathering disguised as low-wage manual work. Indeed, it would be a bit like those human actors who were employed in the early days of generative video’s expansion, not understanding that they were training their ‘synthespian’ AI replacements.

In short, seasonal farm labourers would be training robots to replace them. Ethical complexities aside, this technology means that humans could teleoperate – and train – humanoids in all manner of other industries, too, where safety or access are issues: in hazardous environments such as the space and nuclear sectors, for example.

However, there would be challenges in establishing real-time communications in some environments, especially deep space, given the limitations imposed by the laws of physics: the further away a robot is, the harder it is to control it by radio/light without significant timelag.

On the face of it, this would be an obstacle to remote teleoperation on many farms, too: picking some crops, like soft fruits, is a complex task, and even a millisecond’s delay could damage a fruit. Meanwhile, rural broadband connectivity remains a problem.

That aside, I will make a prediction: it seems likely that in the early days of the domestic humanoid market, some robots in our homes will be the teleoperated avatars of remote workers, who we invite into our homes. This would be an ethical, privacy, and security minefield, of course, but (it seems to me) also a logical inevitability, given the difficulties of training robots in different home environments.

Some robot makers, such as 1X, are putting humanoids on sale or rent this year, focused on complex domestic tasks, even though no robot is currently capable of performing such tasks in unstructured or unpredictable environments unless it is teleoperated. So, draw your own conclusions.

This is the Everest of humanoid robotics: creating a general intelligence, world view, ability to learn, and understanding of the physical environment – an intelligence that is also capable of interpreting verbal commands and putting them into action. In the years ahead, that space will be filled by a mix of cutting-edge technologies and multimodal AI: Visual Language Models (VLMs), Visual Language Action models (VLAs), Large Behaviour Models, visual reasoning, and simulations of the physical world in data.

This is why billions of dollars are being invested in companies such as NVIDIA, FieldAI, DYNA Robotics, Elorian, Physical Intelligence, Luma.AI, Advanced Machine Intelligence Labs, and Jeff Bezos’ Project Prometheus: all are working on cracking the problem of creating a general intelligence for robots that can understand physics, space, and objects, not just language.

But until then? Tele=operation might be the down-and-dirty route to selling a robot that can perform complex tasks in unpredictable environments: just attach a remote human. ‘Only connect’, as the saying goes.

Even so, Extend Robotics’ technology could represent a significant step towards solving one of the most complex challenges in robotics: automating agriculture on a global scale, rather than developing dozens of niche solutions locally.

Hossain says: “We can make farms almost fully automated in the next few years, starting from picking the produce, to transporting the produce to warehouses, then using robots to perform all the different tasks in packaging.”

Ultimately, a single human in the loop could operate fleets of robot workers, he claims, once all the training data had been transferred to each robot:

We can save all the data from the high-skilled work and then use that to automate multiple robots, so the skills become transferable. It’s a huge opportunity.

My take

Perhaps he is right about that. But the complexities and ethical, privacy, and security dimensions of this business model would seem to be off the scale, yet technologists don’t appear to be considering them. That aside, it also seems like a concept that other, more highly capitalised vendors could copy, if they are not already doing so. I guess we will find out soon.