How agricultural robots can revolutionise sustainable farming

Farmers are expected to produce more food, use fewer chemicals, reduce fuel use, protect soil and biodiversity, and manage with less labour. 

These demands make farming more complex. They also make clear that agriculture needs more precise and sustainable ways of working. 

No longer a distant idea

Autonomous robots are increasingly presented as part of the answer. But before we ask what these machines can do, we should ask a more important question: 

How can they be trusted to work safely in real fields, around real people, under real conditions?

That question matters because agricultural robots are no longer a distant idea. They are already being developed for tasks such as seeding, weeding, disease treatment, and harvesting. 

At NMBU, for example, the robotics group has developed the so called 'Thorvald platform,' a modular field robot designed for several agricultural operations. What makes this development important is not only that the platform is autonomous, but that it points toward a different way of thinking about field work. 

Safety first

Instead of relying only on large, heavy machines, agriculture can move toward lighter and more flexible systems that work with greater precision, reduce unnecessary treatment, and place less burden on both workers and the soil.

This is the promise of agricultural robotics. But promise alone is not enough. If these systems are going to be used in vineyards, orchards, tunnels. and open fields, then safety has to come first.

Why we are using robots in agriculture

Robotics has become compelling in agriculture for practical reasons. Farmers are asked to grow more food with fewer resources, use less fuel and fewer chemicals, protect soil and biodiversity, and cope with labour shortages. Automation is about precision, timing, and sustainability.

One useful lesson comes from autonomous cars. They have shown both how impressive autonomy can be and how fragile it becomes in unusual, messy situations. 

Engineers sometimes call this 'the long tail': the huge number of situations that are individually rare but collectively unavoidable. 

A temporary roadwork layout, glare from low sun, snow covering markings, or a child running out from behind a parked car may all be manageable for a human driver, yet difficult for an autonomous system. Safe behaviour depends on perception, prediction, and planning under uncertainty.

That same lesson applies in agriculture. A field robot may perform well in a clean demonstration and still struggle when the environment becomes less tidy. Shadows shift. Dust rises. Plants lean into the path. Soil changes from firm ground to mud or deep ruts. 

What looked safe a moment ago may no longer be safe. In farming, change is not the exception. It is the normal condition.

Risks and challenges in the field, and why these differ from the lab

This is one reason why agriculture can be harder than it appears from the outside. A robot in a lab can look reliable because the world around it has been simplified. 

Lighting is controlled. Obstacles are known. Scenarios can be repeated. Simulation and lab testing are essential parts of engineering, and without them progress would be slow and risky. But they can also create a false sense of confidence. 

The real challenge begins when the machine leaves those controlled settings and enters a living, changing environment.

Sensing isn’t 'seeing' – and AI can be a black box

The difficulty is not only movement. It is also perception. For a robot, sensing is not the same as seeing. It is interpretation under uncertainty. Cameras, lidar, and other sensors do not deliver certainty; they deliver data that software must interpret. 

Machine learning has made this process much more capable, especially in tasks such as detection and classification. But it also introduces a serious safety challenge. 

A model can be confidently wrong. It can degrade when weather, lighting, or crop conditions change. And when something goes wrong, it is often difficult to explain why.

That matters because people working near a robot should never have to guess what it is about to do in order to stay safe.

So the real engineering question is not whether robots can be made impressive. It is whether they can be made trustworthy. That requires a more honest starting point: sensing will never be perfect, models will always have limits, and human behaviour around the machine will remain partly unpredictable. 

In that setting, safety cannot depend on optimism, polished demonstrations, or large operating-hour numbers. It has to depend on assurance.

At NMBU, safety is part of the robot from the start

If that is the lesson from autonomous cars, the next question is what a more responsible path looks like in agriculture.

At NMBU, safety is built into the system from the beginning, and the robot is tested in stages before it is trusted in real field work. It is not just something we check at the end.

The methodological idea is to treat safety as part of the system design, not as a final check. The robot is developed through staged testing: first in simulation, then in the lab, and only then in controlled field conditions.

Formal methods play a role in that process. Key parts of the safety controller can be modelled and checked before deployment, mathematically and systematically, so important safety properties are verified early rather than assumed. 

Runtime verification then carries that assurance into the field by checking whether safety requirements are still being satisfied during operation.

We also use predictive monitoring, which estimates whether the robot is approaching a state where a safety rule may soon become impossible to satisfy. That gives the system a chance to act early.

Why this matters to society

Public trust in robotics will not be earned by the most polished demonstration. It will be earned when systems remain understandable in messy conditions, when their limits are taken seriously, and when safety is treated as a design requirement rather than a public relations afterthought. 

Agriculture is an especially important test case because it is a field where autonomous systems could bring real benefits while still being developed with caution, transparency, and evidence.

Safety is often described as the thing that slows innovation down. In reality, it is what makes innovation usable. A robot that works beautifully in ideal conditions but fails unpredictably in the field is not progress. A robot whose behaviour can be checked, explained, and constrained is.

If autonomous systems are going to become part of everyday agricultural work, they must be safe before they are smart. That is not a barrier to progress. It is the only way progress will deserve trust.

Read also:

Is technology replacing human care?

Is taking control over our own data wishful thinking?