Artificial intelligence is rapidly redefining the capabilities of industrial robotics. While robots have long been used for repetitive automation tasks, the integration of AI is allowing machines to perceive their environment, adapt to change, and make intelligent decisions in real manufacturing environments.
Agile Robots, a robotics company originally spun out of the German Aerospace Center (DLR), is developing systems that combine advanced robotics with artificial intelligence to create adaptive automation capable of operating in complex industrial environments.
In this interview, Sven Parusel, Head of Research Partnerships at Agile Robots, explains how Physical AI is reshaping robotics automation, why industrial data is critical for scaling intelligent robots, and how AI-driven robotics could transform the role of engineers and technicians across modern factories. Agile Robots Physical AI robotics represents a new generation of intelligent automation systems designed to operate in dynamic manufacturing environments.
Agile Robots describes itself as combining robotics with advanced AI. What distinguishes your approach from traditional industrial robotics automation?
Sven Parusel, Head of Research Partnerships at Agile Robots, explains:
At Agile Robots, we leverage the power of Physical AI. Instead of relying on rigid instructions, robots gain the ability to perceive, interpret, and respond to their environment in real time. They no longer just execute predefined paths – they adapt, optimize, and make decisions based on what is actually happening around them.
Physical AI transforms robots from automated tools into adaptive systems – capable of handling variability, learning from experience, and operating effectively in dynamic, unpredictable environments.
While many competitors rely solely on synthetic or generic data, Agile Robots also trains its AI models with real data from our own manufacturing. Combining real industrial data with simulated and human-generated data enables robots to rapidly adapt to new tasks, and to execute them with high precision.
Since our founding as a spin-off from the German Aerospace Center (DLR), we have developed, designed and manufactured nearly all our products in-house. As a result, we have developed a valuable repository of data that has proved to be a strategic advantage.
Industrial robots have existed for decades. Where does AI genuinely change the capability of a robot rather than simply improving programming efficiency?
Sven Parusel:
Robots without AI are merely machines capable of movement but not intelligence. Traditional programming always works the same way. You define a position and a trajectory, and the robot simply follows it. When conditions change, the process must be manually adjusted.
AI elevates a robot’s capabilities beyond the limitations of traditional robotics, moving it from fixed, preprogrammed actions to adaptive intelligence. Traditional computer vision has long allowed robots to react to changes in their environment, but each solution typically required expert knowledge to design a custom system for the specific task.
With Physical AI, robots can perceive, interpret, and respond to their environment in real time, handling variability and uncertainty that conventional robots cannot.
In this way, AI does more than make programming more efficient. It transforms robots into adaptive systems that learn from experience, optimize their actions, and operate effectively in dynamic, real-world conditions that traditional industrial robots cannot handle.
Where are customers seeing the most immediate measurable returns from AI-enabled robotic systems, whether in cycle time, defect reduction, labour efficiency, or throughput?
Sven Parusel:
AI-enabled systems offer measurable results in all of the above.
By adapting in real time, AI-driven robots reduce delays caused by part misalignment, inconsistent materials, or workflow changes, accelerating cycle times without sacrificing quality.
At the same time, these robots handle repetitive or complex tasks autonomously, allowing human workers to focus on higher-value activities.
Adaptive AI systems also maintain consistent performance in dynamic environments, enabling faster changeovers and higher throughput for small-batch or customized production.
In short, AI transforms production from rigid, step-by-step operations into intelligent, self-correcting workflows that deliver immediate, measurable gains in speed, quality, and efficiency.
Traditional robots require precise programming. To what extent can your systems learn tasks, adapt to variation, or improve performance over time?
Sven Parusel:
Our AI-driven automation solutions leverage our own Robotics Foundation Models. They are specifically designed to integrate multimodal inputs such as camera images, tactile measurements, and human commands to break down complex tasks into actionable motion sequences and interactions in the physical world.
Overall, generative AI models form the cognitive-control core of our Physical AI systems. They allow our robots not only to process data but to understand meaning and context and translate this understanding into adaptive physical behaviour.
Through methods like reinforcement learning and imitation learning, where the AI gradually learns optimal strategies by interacting with its environment, these systems continuously improve. Combined with sensor fusion, multimodal representation, and generative learning, this creates a new form of embodied intelligence that increasingly blurs the line between the digital and physical worlds.
How important is real-time, on-device AI processing in robotics compared to cloud-based intelligence, especially in high-speed manufacturing environments?
Sven Parusel:
In high-speed manufacturing, real-time movements are critical because even the slightest delays can accumulate and affect overall cycle time. Cloud-based intelligence introduces latency, bandwidth dependency, and potential connectivity risks that are unacceptable in such time-sensitive processes.
That’s why we rely on onboard AI, which ensures ultra-low-latency decision making, deterministic response times, and uninterrupted operational continuity. While cloud-based intelligence is ideal for large-scale model training, it complements rather than replaces on-device processing.
By combining onboard AI with cloud-based insights, we deliver an architecture that provides immediate, precise control on the shop floor while continuously improving performance over time.
There is concern that intelligent robots may replace skilled labour. How will AI-driven robotics reshape the role of technicians and production engineers over the next five years?
Sven Parusel:
This concern is as old as technology itself. Yet history shows that new innovations tend to create opportunities rather than eliminate them.
AI-driven robotics will transform the role of technicians and production engineers, shifting their focus from manual operation to oversight, optimization, and innovation. Rather than performing repetitive or highly precise tasks themselves, these professionals will increasingly supervise intelligent systems, analyze performance data, and fine-tune AI-driven processes.
In short, humans will focus on higher-value work such as problem solving, process improvement, quality assurance, and innovation while robots handle repetitive, hazardous, or complex physical operations.
As factories move toward Manufacturing-X and connected data ecosystems, how do AI-driven robotic systems integrate with MES, digital twins, and broader production data platforms?
Sven Parusel:
AI-driven robotic systems are no longer standalone solutions but connected, intelligent components within a Manufacturing-X ecosystem.
Integrated with MES these robots can receive dynamic work orders, adjust execution in real time, and return detailed production and quality data for traceability and closed-loop optimization. AI capabilities including LLM-based code generation make adapting to any MES API straightforward and simplify integration across diverse factory environments.
Through digital twins, robots are virtually commissioned, continuously synchronized with simulation models, and optimized using predictive analytics and AI training environments.
Powered by Robotics Foundation Models trained on real-world, synthetic, and human demonstration data, these systems can understand complex environments, adapt to variability, and continuously improve. As a result, robotics becomes a fully integrated, data-driven learning element of the broader connected production platform.
Deploying a robot cell is one thing. Scaling intelligent robotics across global production networks is another. What are the biggest barriers to scale today?
Sven Parusel:
In AI-driven robotics, the barrier is always the data.
Our Robotics Foundation Models enable advanced perception and decision making, but they depend on large-scale, diverse datasets to perform reliably.
Unlike language models trained on internet-scale text, robotics models require multimodal, embodied data including vision, force, motion trajectories, task context, failure cases, and human corrections. That data is far harder to collect and standardize.
The core problem is insufficient scale and diversity. Most industrial datasets are narrow, task-specific, and proprietary. They do not capture enough variability in objects, environments, edge cases, and process drift to support true generalization.
That is why we continue to invest in large-scale data collection by combining real-world factory data with synthetic simulations and human demonstrations to ensure our models can scale reliably across diverse production environments.
As you look toward 2027, what developments in AI and robotics will most significantly impact European manufacturing competitiveness?
Sven Parusel:
Robots without AI are simply machines following fixed instructions that fail when conditions change. At Agile Robots, we leverage Physical AI, giving robots the ability to perceive, interpret, and respond to their environment in real time.
Instead of executing predefined paths, they adapt, optimize, and make decisions dynamically. Physical AI will continue to transform global manufacturing by enabling adaptive, efficient, and resilient production capable of handling variability and learning from experience.
This leads to faster production, higher quality, lower costs, improved safety, and greater flexibility to meet changing market demands.
MTN Analysis: Why Physical AI Could Redefine Industrial Robotics
The interview with Sven Parusel highlights a significant shift underway in industrial robotics. Traditional robots have been highly effective in structured production environments, particularly in high-volume manufacturing where tasks remain consistent. However, these systems often struggle when production requires flexibility or when parts and materials vary.
Physical AI aims to address this limitation by enabling robots to perceive their surroundings and adjust their actions in real time. Instead of relying purely on fixed instructions, robots can interpret sensor data, learn from experience, and respond to changing production conditions.
A key challenge discussed in the interview is the role of data. Robotics AI requires large volumes of multimodal data that include vision inputs, force feedback, motion trajectories, and environmental context. Collecting and structuring this type of industrial data remains one of the biggest barriers to scaling intelligent robotics across factories.
Another important factor is computing architecture. While cloud infrastructure is valuable for training large AI models, real-time robotics requires ultra-low latency processing directly on the robot. Many advanced systems now combine on-device intelligence with cloud-based training environments. This shift toward intelligent automation is part of a broader trend across modern factories where manufacturers are investing heavily in AI-driven manufacturing technologies to improve productivity, quality control, and operational flexibility.
For manufacturers, the long-term implications are considerable. If Physical AI continues to mature, robotics will become capable of handling far more complex tasks, including high-mix production, small-batch manufacturing, and precision assembly operations that traditionally relied on human skill.
FAQ: Agile Robots and Physical AI in Manufacturing
What is Physical AI in robotics?
Physical AI refers to artificial intelligence systems that allow robots to perceive and interact with the physical world. Instead of following fixed instructions, robots can interpret sensor data and adjust their actions dynamically.
How does AI improve industrial robotics?
AI allows robots to detect variation, learn from data, and optimize their behaviour. This enables them to operate effectively in manufacturing environments where production conditions frequently change.
What are Robotics Foundation Models?
Robotics Foundation Models are large AI models trained on multimodal industrial data such as vision inputs, motion trajectories, tactile measurements, and human demonstrations. These models allow robots to perform complex tasks and improve performance over time.
Why is on-device AI important for robotics?
Manufacturing processes require extremely fast response times. Running AI directly on the robot ensures low latency and avoids delays caused by network connections or cloud processing.
What is the biggest challenge for scaling AI robotics?
The biggest challenge is collecting large and diverse datasets. Robotics systems require data that includes physical interaction with objects and environments, which is more difficult to gather than traditional digital datasets.
Will AI robots replace factory workers?
Most experts expect AI-driven robotics to change job roles rather than eliminate them. Engineers and technicians will increasingly supervise intelligent systems, analyse production data, and focus on process optimisation.
