VAIVA Insights

The following video shows our physical test vehicle bearing a development version of the autonomy kit driving autonomously towards operator-defined waypoints while dynamically avoiding obstacles along the way.

What it doesn't show is the thousands of decisions being made every second beneath the surface.

Before a vehicle with our autonomy kit can reach a destination on its own, it first has to answer three fundamental questions: Where am I? What is around me? And how should I get there?

We asked three members of our development team exactly those questions.

How does a vehicle know where it is?

Anjaly: "The honest answer is that a vehicle never knows with absolute certainty. Every sensor has limitations. Wheel odometry drifts over time. IMUs accumulate errors. GPS is not always available or accurate enough.

No single sensor provides a complete or sufficiently reliable picture on its own. Our job is to continuously fuse multiple imperfect sources of information into the best possible estimate of the vehicle's position."

What's unusual about our current physical demonstrator is that we deliberately use significantly simpler sensors than we will use on a future operational UGV.

Anjaly: "That creates additional challenges, but that's exactly why we do it. If we can make the system work reliably with lower-cost sensors, we're forced to solve many of the difficult localization problems early. Better sensors can improve performance later, but we don't want to depend on them."

How does a vehicle understand its environment?

Tim: "Knowing where you are is only part of the story. You also need to understand what's around you. As the vehicle moves, it continuously builds a map of its surroundings using LiDAR and its current position estimate.

At the same time, that map is used to validate its assumption of motion and corrects the vehicle's own localization. This feedback loop is what makes SLAM so powerful. The vehicle learns about the environment while the environment helps the vehicle determine where it is."

The team sees this as another benefit of testing with simpler sensors.

Tim: "When your sensor data is less than perfect, you quickly discover which assumptions hold up in the real world and which ones don't. Those lessons are incredibly valuable before moving to larger and more capable platforms."

How does a vehicle decide where to drive?

Once the vehicle has an understanding of its position and surroundings, it can start making decisions. The operator only specifies a destination. Finding a safe path is the vehicle's responsibility.

Shipra: "The interesting question is usually not whether a path exists. It's which path is the right trade-off between safety margin, progress and vehicle constraints. The planner evaluates obstacles, drivable free space and alternative routes continuously, and updates its solution as the environment estimate changes.

We don't see autonomy as following a predefined line. The vehicle only commits to the next short segment, then reassesses the situation and decides what to do next."

That flexibility in decision making is one reason why key planning components are developed in-house.

Shipra: "We want the freedom to consider additional factors in the future, such as terrain conditions, vehicle characteristics or mission-specific objectives."

Looking Ahead

The current demonstrator already combines localization, mapping and autonomous waypoint navigation into a single system. The next milestones will bring those capabilities into larger operating areas and increasingly realistic environments.

For the localization team, the next step is clear:

Anjaly: "We're looking forward to adding GPS-based georeferencing. That will allow us to move beyond local navigation and start operating over much larger distances."

For the mapping team, the focus shifts from building maps to validating them under more demanding conditions:

Tim: "Real off-road environments are unpredictable. Testing in larger and more diverse terrain will help us understand how to improve our localization and mapping estimates."

And for the planning team, those tests are where everything comes together:

Shipra: "The most exciting part is seeing the complete system operate in real off-road scenarios. That's where navigation, perception and planning all have to work together as one."

Because in the end, what makes autonomy interesting is not the driving itself.

It's everything the video doesn't show.

Wenn Sie sich ein autonomes Fahrzeug vorstellen, stellen Sie sich wahrscheinlich vor, wie es sieht – wie es das Gelände analysiert, Hindernisse erkennt, einen Weg findet. Sie denken wahrscheinlich an farbige Overlays, die befahrbare Regionen, Hindernisse und Geländemerkmale hervorheben und den Eindruck erwecken, dass die autonome Navigation weitgehend gelöst ist. Aber das Sehen ist die leichtere Hälfte. Die eigentliche Arbeit beginnt in dem Moment, in dem sich das Fahrzeug entscheiden muss.

Letzte Woche haben wir in From Data to Dirt an einem Beispiel beschrieben, wie unsere Wahrnehmungsalgorithmen lernen, befahrbaren Boden von allem anderen zu unterscheiden. Das Ergebnis ist eine Karte des befahrbaren Bereichs, in dem eine Bewegung möglich ist. Auf einer asphaltierten Straße beantwortet diese Karte die Frage fast von selbst. Im Gelände ist das nicht der Fall – denn "möglich" und "sinnvoll" sind zwei völlig verschiedene Dinge.

Wenn Sie ein Feld oder einen Waldweg überqueren, finden Sie in der Regel nicht nur einen Weg vor, sondern mehrere. Einer ist kürzer, aber schlammig. Ein anderer bietet bessere Traktion, führt aber über einen steilen Hang. Ein dritter ist weder für die Traktion noch für die Steigung ideal, scheint aber ein guter Kompromiss zu sein. Die Frage lautet also nicht mehr "Kann ich hier fahren?", sondern "Welchen dieser Wege soll ich nehmen – und bringt er mich dorthin, wo der Auftrag mich hinführen soll?"

Die Beantwortung dieser Frage ist die Aufgabe der Trajektorienplanung. Und sie geht immer von der Absicht aus: dem Missionsziel und den Wegpunkten, die der Bediener festgelegt hat. Ein einfacher logistischer Abwurf kann den Schwerpunkt verändern und damit auch die Steigungen, die ein Fahrzeug auf dem Rückweg bewältigen kann. Die Mission macht zusammen mit dem definierten Ziel, den gewünschten Wegpunkten und der aktuellen Position des Fahrzeugs aus einer offenen Karte eine reale Aufgabe – gelange von hier nach dort, durch dieses Gelände und berücksichtige diese Einschränkungen.

Im Rahmen dieser Aufgabe jongliert der Planer mit konkurrierenden Anforderungen auf einmal – Steigung, Traktion, Hindernisse und die einfache Frage, ob das Fahrzeug die Bewegung physisch bewältigen kann. Selten weisen sie in dieselbe Richtung: Die kürzeste Linie ist nicht die sicherste, die sicherste nicht die schnellste. Der Planer wägt diese Faktoren immer wieder gegeneinander ab und entscheidet sich für eine Trajektorie, die nicht nur auf dem Papier möglich ist, sondern auch im Dreck funktioniert – eine, der das Fahrzeug dann selbstständig folgen kann.

Genau dieser Balanceakt ist es, den die Autonomie im Gelände erfordert, denn es gibt keine vorgezeichnete Linie, der man folgen kann. Das Fahrzeug muss seine Umgebung erfassen, die Optionen mit dem Auftrag abgleichen und sich entscheiden. Die Wahrnehmung sagt ihm, wie die Welt aussieht. Die Planung sagt ihm, wie es sich durch diese Welt auf ein Ziel zubewegen soll. Erst zusammen ergeben sie eine autonome Navigation.

Denn bei der Autonomie geht es nicht wirklich um das Sehen. Es geht darum, Entscheidungen zu treffen.

From Data to Dirt: How We Develop Off-Road Autonomy

When people picture an autonomous vehicle, they usually imagine the finished product: a machine moving through rough terrain, no driver, no hesitation. What stays invisible is everything that happens long before that first autonomous drive across a field or forest track — the slow, deliberate work of teaching software to understand the world around it.

For us, that work begins with data.

Every capability begins with observation

Before the Autonomy Kit can make a single decision, it has to understand what it sees. Is this surface drivable? Does the path continue around the corner? Is that vegetation an obstacle — or just scenery?

Answering these questions takes real-world observation. So we start by recording sensor data in representative environments. Every route, every terrain type, every change in weather and light adds another piece to our understanding of the operational domain.

The principle is simple: collect real data first, instead of trying to solve every imaginable problem on paper.

Turning data into insight

Raw recordings are where the engineering really starts. We use them to test early versions of our perception software and to compare different approaches under identical conditions.

This is how theory becomes knowledge. Patterns become visible, challenges become concrete, and assumptions turn into hypotheses we can actually test. Step by step, a clearer picture emerges of how an autonomy function should behave in the field.

From insight to virtual validation

Once we've got those recordings, the real work begins. We use them to test early versions of the perception software and to put different approaches head-to-head under exactly the same conditions. That last part matters more than it sounds — when everything runs against the same data, you can actually tell which approach is better, instead of guessing.

This is the stage where raw recordings turn into genuine engineering knowledge. Patterns start to jump out. The tricky edge cases become visible. And vague ideas turn into concrete questions we can actually test. By the end of it, we have a much clearer sense of how a given function should behave out in the field.

Closing the loop with physical demonstrators

Of course, virtual validation alone is never enough. At the end of the day, the Autonomy Kit has to work in the dirt.

That's why we bring in physical demonstrators early and step by step, instead of all at once. Small platforms and prototypes let us validate integration, perception behaviour and system interactions before we move to the larger UGVs — our actual target platform. It's the best of both worlds: the speed of software development, grounded by the honesty only real hardware can provide.

Why this matters

The real value of this process isn't just efficiency — it's confidence. Every stage adds evidence. Every validation reduces uncertainty.

By the time a feature reaches the target platform, it has already passed through multiple learning loops: software development, data collection, virtual analysis and physical testing. So we don't treat UGV trials as the starting line. They're simply the next step in a process that's been running all along.

The path from data to dirt

Off-road autonomy comes down to one hard problem: operating reliably in a world that's messy and unpredictable. Pulling that off takes more than a clever vehicle. It takes data, analysis, understanding and testing — and, just as importantly, a way of working that ties all four together.

That's what "from data to dirt" really means for us. Not a single project, but a loop we keep running: start by observing, end up in real terrain, and let every step build toward autonomy you can actually trust.

An integral part of the new Vehicle Tech Week, Autonomous Vehicle Tech Expo, Europe returns to Stuttgart, Germany, on June 23–25, 2026, bringing together the latest autonomous and ADAS technologies, development tools, and next-generation mobility solutions. For VAIVA, this is one of the most important events of 2026. Matthias Haringer, our Principal Virtual Development lead, will share the latest information on ICASIS, the platform we are establishing as a reference for in-cabin simulation. We will also have a booth there and many great colleagues on-site. Stop by to get all the details, use cases, and roadmap updates.

InCabin.Sensing brings together over 400 of the most influential practitioners, technical and HMI experts, OEMs, Tier 1s, and research institutes to discuss the latest technical advancements, market trends, and opportunities. VAIVA will showcase ICASIS feature updates, as well as a sneak peek into what is next on the roadmap. Drop by and connect with Florian to learn more about our In-Cabin simulation platform.