No more exaggerated promises: That’s why robotics needs its own levels of autonomy

When it comes to autonomous driving, it is clear what we are talking about when it comes to autonomy levels. Level 2 means assistance, level 5 means complete autonomy. However, such a common language is still missing in the development of autonomous industrial robots.

This leads to a fundamental problem: the term “autonomous robot” is used to refer to many completely different systems. Sometimes this means a classic industrial robot, sometimes a system with cameras, sometimes a machine that learns from data. There are big technical differences between these approaches, but often not linguistically.

This has concrete consequences and causes confusion. Companies don’t know exactly what they are buying. Expectations are set too high or even too low. And development teams work with different ideas about what autonomy actually means. This slows down progress in the industry because there are no clear goals.

A uniform orientation is needed that makes these differences tangible. A classification could help – for example in the form of a five-level scale, similar to autonomous driving.

Robots: Five stages on the path to autonomy

Level 1: Fixed processes

A robot works strictly according to plan. Every movement is predetermined, but is optimized using a digital twin of the robot and fixed environmental geometries. Example: A robot always grabs a component in the same place and places it in the same place by using a path planner to optimize the robot movement between a start and destination point – even around obstacles if these were previously included in the digital twin. As long as everything fits exactly, it works reliably. If the part shifts, the process fails.

Level 2: See and react

The robot can perceive its surroundings and react to them. Example: Parts are unsorted in a box. A camera detects your position. The robot adjusts its grip and still lifts them out correctly. He no longer blindly follows a script, but reacts to the real situation. Many applications in logistics and manufacturing today operate precisely on these two levels.

Level 3: Learning individual skills

Certain partial movements are no longer programmed, but rather learned. Example: A plug must be inserted precisely into a socket. Small deviations quickly lead to errors. Instead of specifying every movement, the robot learns through demonstration how to insert the plug correctly and compensate for small inaccuracies itself. These movements are embedded in larger processes, such as gripping the plug. This is often the point at which classic automation reaches its limits and learning-based approaches begin.

Level 4: Complete tasks in a familiar environment

The robot can carry out complete processes independently and using fully learned behavior as long as the environment is familiar. Example: A robot takes over the loading and unloading of a machine. The parts vary slightly in shape and position. Nevertheless, he recognizes them, grasps them correctly and carries out the entire process independently. The process is no longer just a sequence of individual movements, but rather a coherent process. However, if the robot is entrusted with a new task or the environment changes significantly, it must be retrained.

Level 5: Solve new tasks yourself

The robot can understand new tasks and implement them independently. Example: A system receives the order to sort delivered parts for the next production step. It recognizes the objects, decides on the order and adapts its behavior, even if the exact situation has not been trained beforehand. However, the same robot can also be used to put parts together without having to be retrained at great expense – it responds flexibly to speech and a few demonstrations. Only here can we speak of general autonomy in the narrower sense.

Why this classification is important

The differences between each level are not academic but practical. A system at level 2 can react to deviations, while a system at level 4 can take over entire processes independently. However, both are now often referred to as autonomous, even though their capabilities are fundamentally different.

If clear terms are missing, misunderstandings arise. Companies expect flexible, robust systems and receive solutions that only work reliably under stable conditions. At the same time, real progress is underestimated because new capabilities are not precisely identified.

A uniform scale provides orientation here. It makes it clear that autonomy does not arise suddenly, but rather develops step by step. The individual skill levels become clearly distinguishable and their respective added value becomes tangible.

Physical AI: What already works today and what is still missing

From level 3 onwards, algorithmic processes are increasingly replaced by completely learned ones. They require the use of neural networks that have increasingly internalized the characteristics of the physical world – in robotics we speak of “physical AI”: i.e. understanding the geometry, physics and material properties of the environment. This area of ​​development has existed in research for around ten years, but has seen major investments in research and industry since 2023 due to the success of large language models that use similar learning methods.

Levels 1 and 2 have long been established in many industries. Robots work with high precision, recognize objects and react to minor deviations. The first applications of Level 3 and Level 4 are already in use, for example for sensitive movements or more complex gripping processes.

What robotics is still missing is true generalization – i.e. level 5. Systems that can master new tasks flexibly and without special training are still a thing of the future. This is precisely why a clear classification is crucial. It prevents existing systems from being overestimated and at the same time shows where substantial progress is actually taking place.

More clarity about the status of development

Robotics is developing quickly, but not in leaps and bounds. Progress occurs along clear developmental steps, from stable, deterministic processes to perception and learning-based skills to generalization.

As long as all these stages are summarized under the same term, it remains unclear what is actually being discussed. Therefore a common language is needed. This leads to more realistic expectations, measurable progress and a more objective discussion. The crucial question is not whether a robot is autonomous, but to what level.