Robotics development on open source Python platforms, our new business area for 2027
Robots now learn from demonstrations and understand language. Why Anexia Digital Engineering is building robotics development with ROS 2 and LeRobot into a new business area for 2027.

When robots learn
For decades, a robot was a machine that did exactly what a specialist had spent weeks programming into it. That is now changing fundamentally. Robots learn tasks from demonstrations, understand instructions in natural language and cope with situations nobody anticipated. The same methods that made language models big are behind this shift. For a company whose strength lies in artificial intelligence and software, it opens up a field that used to belong to traditional mechanical engineering. That is why Anexia Digital Engineering is building robotics development into a new business area, on the foundation of open Python platforms.
“Robotics is strategically important to us and one of the cornerstones for 2027,” says Patrick Kulterer, CEO of Anexia Digital Engineering GmbH. “It sits alongside AI knowledge platforms, agentic enterprise systems operated with digital sovereignty, autonomous driving and AI-assisted procurement. All five fields come down to the same question, namely how artificial intelligence turns into reliable work.”
Why robotics is changing now
Three developments have come together in a short space of time, and they have noticeably changed robotics development.
The first concerns the models. Research groups and companies have presented methods that translate camera images and language instructions directly into a robot's movements. These models are trained on demonstrations in which a person operates the robot remotely. Instead of programming every movement, you show it.
The second concerns the data. For a long time, robotics lacked what the internet had been for language understanding, namely a large, open collection of examples. That collection is now growing. According to a publication by the LeRobot developers, the Hugging Face platform alone held more than 16,000 datasets from over 2,200 contributors in September 2025.
The third concerns cost. Robot arms that anyone can rebuild from openly available designs now cost a fraction of what laboratory equipment used to. Simulations run on graphics cards found in every data centre. With modest effort, a team can try out what would have required a research budget only a few years ago.
Taken together, the bottleneck is shifting. It lies less and less in mechanics and more and more in software, data and the operation of AI models. That is why robotics development is now within reach for a software company.
Why Python and why open source
Two fundamental decisions shape our robotics development, the choice of programming language and the openness of the platforms.
Python is the language of AI
Almost every relevant tool for machine learning is driven from Python. Models are trained with PyTorch, data is prepared with proven libraries, and experiments are described in a few lines. Anyone doing robotics development with learned methods inevitably works in this world.
Then there is the pace. In Python, an idea can be tried out in an afternoon. In robotics development, which changes month by month, that speed is worth more than the last microsecond of compute time.
One limitation belongs here too. Control loops that run a thousand times per second, and anything that concerns human safety, do not belong in Python. That is where C++ and certified controllers have their place. Good robotics development combines both worlds, with Python for perception, planning and learning, and hardware-level technology for everything that has to happen in hard real time.
Open platforms instead of closed systems
Classic industrial robots are tied to their manufacturer's software. Switch brands and you start programming all over again. Open platforms remove this lock-in and make robotics development independent of any single manufacturer.
For us and our clients, this has tangible advantages. The source code is open to inspection, which is what makes security and quality reviews possible in the first place. A worldwide community from research and industry contributes to development, so bugs are found quickly. Knowledge and code carry over to the next robot type. Licence costs for the foundations disappear, and the budget goes into the actual solution. And European companies stay independent of individual vendors and their pricing policies.
That last point deserves weight. Digital sovereignty does not end at the data centre. A company that makes its production dependent on a closed platform whose terms can change at any time has a strategic risk on its own premises.
The platforms we rely on
Two tools form the core of our robotics development, and others complement them depending on the task.
ROS 2 as the backbone
Despite its name, the Robot Operating System is not an operating system. It is a collection of building blocks through which a robot's sensors, drives and programs exchange messages. For more than a decade it has been the standard in research, and increasingly in industry. Support for the first generation ended in May 2025, and today only ROS 2 counts.
In May 2026, Lyrical Luth was released as a new version with long-term support until 2031. For projects starting now, it is the obvious foundation, provided all the required packages are already available for it. Otherwise, Jazzy, supported until May 2029, remains a solid choice. Anyone still working on Humble should plan the migration, because its support ends in May 2027.
ROS 2 can be programmed in Python and in C++, and both languages work together in the same system. That is exactly what makes it such a good fit for our robotics development. For navigation, motion planning and coordinating entire fleets, mature open packages build on this foundation.
LeRobot for learning from demonstrations
LeRobot is an open library from Hugging Face, written in Python and built on PyTorch. It covers the whole path, including motor control, recording demonstrations, and training and running the models. A unified dataset format makes recordings from different robots comparable.
It contains reference implementations of the most important learning methods, ranging between compact models with a few dozen million parameters and large models with several billion. This is valuable in practice, because several methods can be tested against each other on one task without major rework.
In our robotics development, we use LeRobot to try out capabilities quickly. Before a client invests in an installation, a low-cost setup can show whether a task can be solved with learned methods at all, and how many demonstrations it needs.
Simulation and the wider landscape
Training robots in the real world is slow and wears out material. That is why a large share of robotics development takes place in simulation. The market offers several open tools here, including MuJoCo from Google DeepMind, Isaac Lab from NVIDIA and Gazebo from the ROS ecosystem. Which one fits depends on the task, and we decide that for each project.
For autonomous driving, Autoware offers an open software stack built on ROS 2, and CARLA a simulator that can be controlled from Python. The kinship is no coincidence. A vehicle that moves through traffic on its own is, technically speaking, a robot on wheels.
What a software company brings to robotics development
It is a fair question. Why should a company known for software, apps and cloud get robots to work? The answer lies in what makes the difference in robotics development today.
Experience with AI in production
Getting a model to run in the lab is the smaller part. The hard part comes afterwards. Data has to be versioned, training runs documented traceably, models monitored and updates rolled out without downtime. This discipline grew up in the software world and is still the exception in many robotics projects. We bring it with us. Our case study on the video analysis app shows what this looks like with moving images.
“We are very strong in AI,” says Patrick Kulterer. “That strength does not stop at the screen. A robot that learns from demonstrations is, at its core, an AI system with motors. What we have learned about data, models and reliable operations applies there just as much.”
Software craftsmanship
Automated tests, version control, a second person reviewing every change and reproducible builds are everyday practice in software development. In robotics development, this craft decides whether a prototype ever makes it into continuous operation.
Connecting to enterprise systems
A robot that moves boxes is worth little on its own. Value is created when it knows which box belongs where, when it reports to the warehouse management system what it has done, and when a fault lands as a ticket with the right team. Integration with ERP, production control and warehouse management is classic work for a software company, and it is regularly underestimated in robotics development.
Infrastructure and digital sovereignty
In robotics development, training models requires computing power, running them requires reliable infrastructure, and the data generated along the way is often a trade secret. Camera footage from a production line shows products, processes and people. The Anexia Group operates its own data centres in Europe. Our clients' models and data can stay there or in their own environment. We follow the same principle with our hospital billing system, where patient data is involved.
What we offer
Building a new business area means being honest about where you stand. We are not a robot manufacturer and do not intend to become one. Our contribution is the software that turns existing hardware into a solution. Our offering in robotics development is structured in five building blocks.
Assessing feasibility
It starts with a simple question. Can this task be automated reliably with today's means? We answer it with a limited trial on real hardware or in simulation, within a few weeks and with a clear result. If the answer is no, that is valuable information too.
Capturing and preparing data
Learned methods stand or fall with their data. We set up teleoperation, record demonstrations, check their quality and bring them into a format that can be reused. This work is unspectacular, and in robotics development it decides success.
Training and evaluating models
Which method suits the task, how many demonstrations are needed, how does the model behave when conditions change? We train, compare and measure. The outcome is a success rate under realistic conditions, not a video of the one run that worked.
Integrating
The solution is integrated into the installation with ROS 2, connected to sensors and grippers and linked to the company's systems. For safety-related functions, we bring in partners from mechanical engineering and automation.
Operating and improving
A robot in service continuously generates new data. Every failed grasp is a hint where the model can improve. We monitor operations, collect these cases and roll out improved models in a controlled way.
Where we see applications
Not every task suits robots with learning methods. The best fit is work that is too variable for classic programming and too monotonous for people.
This includes gripping and sorting parts that arrive unordered, and loading and unloading machines when products change frequently. Order picking in warehouses with a large and changing range fits just as well, as does handling samples and containers in the laboratory. Further cases are visual inspection where a camera has to be guided, and transport within buildings on mobile platforms.
What these cases have in common is that the environment differs a little every day. That is exactly what rigidly programmed installations struggle with, and exactly where learning methods play to their strengths.
What is not yet possible
An honest article about robotics development has to look at the limits. The videos circulating online usually show the best of many attempts.
Learned methods do not work without errors. A success rate of ninety per cent sounds good, and in manufacturing it means that every tenth grasp misses. Many tasks therefore need checking steps, a second attempt or a person who takes over the exceptions.
Speed is often below that of a person or a conventionally programmed installation. The advantage of learning robots is their flexibility, not their cycle time.
The transfer of what works in simulation into reality does not always succeed. Friction, light and material behaviour are hard to reproduce.
And finally, robotics development is moving so fast that a decision made today can be outdated within a year. That is one more reason why open platforms are the right choice, because they allow the model to be swapped without rebuilding everything else.
Safety and regulation
Where machines move, people's physical safety is at stake, and that shapes all robotics development. This section does not replace legal or safety engineering advice, but there are some dates decision-makers should know.
From 20 January 2027, the new Machinery Regulation applies in the EU. It replaces the previous directive and, for the first time, explicitly addresses machines with learning or evolving behaviour. Risks must also be considered for the period after placing on the market, and protection against manipulation via digital interfaces becomes mandatory. It is now also clearly defined when a modification counts as substantial and triggers a new conformity assessment.
The AI Act comes on top. For AI systems built into products as safety components, the requirements for high-risk systems were recently postponed to August 2028. How the two sets of rules interact for machinery should be assessed separately for each product.
For our robotics development, both lead to one principle. The safety of a system must never depend on a learned model making the right decision. Protective devices, limits on force and speed, and the emergency stop are implemented with conventional, proven technology. The model works within these limits.
How robotics fits our other focus areas
The five fields we are putting at the centre for 2027 do not sit side by side by accident. They share technology, mindset and often clients as well.
AI knowledge platforms
Companies hold enormous knowledge in documents, manuals and people's heads. Knowledge platforms make it accessible. This matters for robotics development, because work instructions, maintenance plans and quality specifications are exactly the knowledge a machine needs for its task.
Agentic enterprise systems, digitally sovereign
Software agents plan tasks, call systems and complete processes across several steps. In this picture, a robot is an agent with a body. The same layer that controls a digital process can trigger a physical one. What matters to us is that these systems run in Europe and under the client's control.
Autonomous driving
Perception, planning, simulation and handling sensor data are the same disciplines. The open tools overlap heavily too. What we learn in robotics development benefits this field directly, and vice versa.
AI-assisted procurement
At first glance, this is the most distant focus area. At second glance, the circle closes. A production line where robots work flexibly needs procurement that reacts just as quickly. Together, they make a supply chain that adapts to changing demand.
“We don't see these topics as separate products,” says Patrick Kulterer. “An agent that finds the right work instruction in a knowledge platform, triggers an order and finally sets a robot in motion is technically within reach. Our goal for 2027 is for chains like these to run reliably for our clients, on infrastructure they control themselves.”
How a project begins
Getting started in robotics development does not have to be a major project. Working in small steps has proven itself in software development, and we apply the same approach to robotics.
First, a single, clearly defined task is chosen. It should occur frequently, take effort today and not cause much damage if something goes wrong.
Next comes the feasibility assessment. The result is a measured success rate and an assessment of what is still missing for continuous operation.
If the result is good, a pilot is set up on the real installation, accompanied by the people who work there. Their experience cannot be replaced, and their acceptance decides success.
Only then is a decision made on scaling up, based on figures from your own operations.
Who should talk to us
Companies in manufacturing, logistics and laboratories that have tasks with high variance and struggle to find staff for them should feel addressed. The same applies to manufacturers of machines and devices who want to extend their products with learning functions and are looking for a software partner. Research institutions that need support with software quality and operations to take the step out of the lab and into real use are addressed as well.
Anyone who wants to automate a task with identical parts in identical positions, on the other hand, is better and more cheaply served by conventional automation. Saying so is also part of honest advice in robotics development.
Frequently asked questions about robotics development
We hear some questions about robotics development in almost every conversation. Here are the answers in brief.
Does Anexia Digital build robots itself?
No. Our robotics development concerns the software, and we work with hardware that is available on the market and with partners from mechanical engineering and automation.
Is open source software suitable for industrial use?
Yes, provided it is used professionally. ROS 2 runs in numerous products. What matters is choosing a version with long-term support, maintaining dependencies carefully and keeping learning functions clearly separate from safety technology.
How many demonstrations does a robot need to learn a task?
Anything between a few dozen and several thousand, depending on the task, the variance and the method. This is precisely the question the feasibility assessment answers.
Do our data and models stay with us?
Yes. Training and operations can take place in your environment or in European data centres of the Anexia Group. Footage from your operations is not used for other clients.
What does getting started cost?
Getting started means a feasibility assessment that is deliberately kept small. We define scope and price after the first conversation, once the task and the available hardware are known.
Let's talk about your task
Which activity in your operations is too variable for a rigid installation and too monotonous for a person? In a no-obligation first conversation, we assess whether and how it can be solved with open software and learning methods. That turns the big topic of robotics development into a concrete first step. Use our contact page to arrange a date.


