How Dutch robotics startups are making robots deployable faster with smarter training methods

A robot that masters a new gripping task in a matter of hours rather than weeks: that is what a growing number of Dutch startups are pursuing with AI-driven training methods. Companies such as Fizyr from Delft and Smart Robotics from Best are combining deep learning, simulation and visual data to lower the barrier to industrial robotisation.

The stakes are high. In logistics, manufacturing and the agricultural sector, demand for automation is rising while the supply of technical personnel is lagging behind. How robots are trained partly determines how quickly and broadly that automation spreads, and which tasks remain for people.

From manual programming to learning systems

Traditionally, setting up an industrial robot required a great deal of manual programming: an engineer had to define rules for every situation. Machine learning is shifting that model. Systems now learn to recognise patterns from large volumes of images or sensor data, and apply what they have learned to new situations.

Fizyr, founded in 2014 as a spin-off of TU Delft, builds deep learning vision software that enables robots to recognise and grasp objects they have never encountered before. The system calculates more than a hundred gripping positions per second and performs immediate quality checks to reject damaged products. Customers such as PostNL, Vanderlande and ABB use the software for sorting and order picking. Fizyr works with supervised learning: algorithms are trained on real-world images so the system learns which approach works for which object.

For standardised tasks such as palletising or sorting known parcels, such systems work quickly. The challenge lies in variation: unknown packaging, changing lighting conditions, irregularly filled boxes. Reinforcement learning can be deployed for those situations, but it demands significant computing time and good simulations.

Simulation as an accelerator

To shorten training time without halting production, several startups use digital twins: virtual replicas of the physical environment in which robots can practise. A robot can run through thousands of scenarios in simulation before it sets foot on the shop floor, reducing time-to-deployment and lowering costs for system integrators.

Avular from Eindhoven, also a TU/e spin-off founded in 2014, applies this to autonomous ground and aerial robots for industrial environments. The company develops its own software for world modelling and sensor fusion, training algorithms in simulation to navigate complex, dynamic spaces. In April 2024, Avular closed a Series A round of €3.5 million.

Generative AI adds another dimension to this. Synthetic training data for rare situations, for example, a robot encountering a damaged parcel or an obstacle in an unusual location, is expensive to collect in the field. Generative models can simulate those situations. The quality of that synthetic data determines how the system behaves in practice later on; it is not a guarantee of robustness, but a way to accelerate and broaden data collection.

Domain-specific applications require customisation

Not all robotics startups focus on logistics. The breadth of the Dutch scene illustrates how strongly the training approach depends on the domain.

Corbotics from Delft is building an autonomous robot for cardiac ultrasound examinations. Whereas a logistics robot learns to grasp objects, the Corbotics robot learns to position an ultrasound probe accurately on the human body. This requires a combination of medical imaging data, haptic feedback and stringent validation requirements. The company received an EIC Accelerator grant of €2.5 million and an Innovation Credit from RVO, with a financial runway through Q1 2028.

Odd.Bot from Rotterdam trains its autonomous weeding robot on agricultural environments: the robot distinguishes crops from weeds based on visual data. With €12 million in Series A funding (2024), the company is scaling up its field operations. Rocsys from Rijswijk combines robotics with computer vision for hands-free charging of electric vehicles and raised the largest funding round in this group at €42.7 million.

Each of these applications requires its own training strategy: the data, the environment and the error tolerances differ substantially. This makes sector-specific knowledge a competitive factor alongside generic ML tooling.

What this means for people on the shop floor

In sectors with structural labour shortages, such as logistics and warehousing, robotic systems are more likely to fill the gap than to displace existing jobs. Smart Robotics from Best, founded in 2015 and having closed a Series A of €8.7 million in 2024, explicitly focuses on order picking and palletising: roles that are difficult to fill due to a lack of personnel. The robots take over repetitive tasks; employees shift towards supervision, maintenance and exception handling.

In the medical sector, the dynamic is different. Corbotics positions its ultrasound robot as support for sonographers, not as a replacement. The robot handles standard image acquisition, allowing specialists to focus on interpretation and patient contact. With an ageing population and growing demand for cardiac diagnostics, that is a relevant perspective for healthcare organisations.

The speed at which robots can learn new tasks also has implications for the flexibility of automation. The shorter the training time, the sooner smaller companies and variable production environments become viable candidates. This makes automation more accessible beyond the large logistics centres and automotive plants where robots have long been commonplace.

For the broader Dutch and European AI scene, the clustering of robotics startups in Delft and Eindhoven illustrates how university spin-offs can build an entire sector. Investors who entered in recent years are now seeing companies grow from seed to Series A with concrete industrial customers. Policymakers focused on digital manufacturing and labour market policy will find that the question is no longer whether robotisation is coming, but how quickly the training infrastructure will scale to match demand from the business community.