First, the current state of play. According to the International Federation of Robotics’ (IFR) World Robotics Report 2026, published on 24 September, 603,000 industrial robots were installed worldwide in 2025, eleven per cent more than in the previous year. The operational stock exceeded five million for the first time. 354,000 units, or 59 per cent, went to China. With 38,400 installations, the US overtook Japan (36,200, down 19 per cent) and is now the second-largest market. Germany ranks fifth with 24,800 units and accounts for 41 per cent of EU installations – which, however, fell by 11 per cent to 60,500. For 2026, the IFR expects an increase of nine per cent to 655,000, and 806,000 units for 2029. The market is growing, just not here.

NVIDIA’s lesson from its own manufacturing

In early October, NVIDIA’s Seattle Robotics Lab set out just how much AI the factory really needs in a remarkably candid report. Working with contract manufacturer Foxconn, the lab was tasked with automating two steps in the assembly of test carriers for the GB300 superchip: inserting a heavy busbar and securing it with 16 screws, and feeding four wired connectors into tight sockets. The production requirements were: a 99.5 per cent success rate, a cycle time no more than double that of a skilled worker, and no collisions. For the busbar, a classic sequence of pose estimation, path planning and impedance control proved so effective that the planned switch to end-to-end learning was dispensed with: over 95 per cent success, 160 seconds instead of the required 124; the bottleneck is the screwing. For the connectors, imitation learning and generic models failed due to too few cables, which were subject to wear and tear; only a specialised pose estimator, 3D-printed gripper fingers and an insertion strategy – pre-trained in simulation and fine-tuned in reality – achieved a success rate of 90 to 95 per cent. The authors call this the ‘reverse bitter lesson’: Where interaction data is scarce and parts are expensive, engineering expertise trumps scalability. And they calculate the cost of reliability: to achieve 99.5 per cent statistical certainty, 598 consecutive error-free attempts are required.

AI is making its way into control systems

Meanwhile, AI is not entering the factory floor in the form of humanoids, but as a software function. On 19 October, KUKA unveiled iiQKA.AI Vision, an integrated 2D and 3D image processing system that runs on an NVIDIA-based processing unit directly within the robot controller and can be configured via drag-and-drop for tasks such as retrieving items from a box, assembly, quality inspection and depalletising – without the need for specialist vision expertise, as the manufacturer emphasises. Teradyne, the parent company of Universal Robots, acquired a stake on 5 October in Bright Machines, a provider of software-defined microfactories with over 130 installations in more than ten countries, which already manufactures AI infrastructure in the US. The reasoning behind both partners’ decision: server racks with short design cycles and frequent revisions cannot cope with rigid automation; they require cells that can be reprogrammed without needing to be redesigned.

A unicorn with a penchant for factories

RobCo demonstrates that capital is following this path. The Munich-based company, founded in 2020, reached a valuation of one billion dollars on 5 October – a doubling within nine months, triggered by a share sale that also provided liquidity for employees. According to its own figures, RobCo has raised a total of around $200 million, including $100 million in January from Lightspeed, Sequoia and Lingotto. The product: modular robotic arms that can be slotted together like building blocks, offered as ‘Robotics-as-a-Service’ for small and medium-sized enterprises. Around 70 per cent of its business is in Europe, whilst the US is growing fastest – which is why co-founder and CEO Roman Hölzl has moved to America and part of the assembly takes place in Austin. On 4 March 2027, RobCo plans to unveil ‘Alfie’ in Munich, a two-armed, self-learning robot designed for tasks that cannot be performed using traditional automation. Sequoia partner Luciana Lixandru explains the decision to invest by stating that AI should not only generate content but also act independently in the physical world. RobCo is deliberately focusing on the factory floor rather than humanoid robots.

Three conclusions for decision-makers

Firstly: the growth in industrial robotics is real, but is shifting towards Asia and the US; Europe’s decline is a result of investment decisions, not market conditions. Secondly: AI is making its mark through control systems and simulation tools, not through new robot designs – and the best results continue to come from combining machine learning with sound engineering. Thirdly: Those who take the ‘reverse bitter lesson’ seriously will recognise it as a locational advantage. Process knowledge, fixture design and control engineering are precisely the disciplines in which German automation specialists have traditionally excelled. Physical AI does not replace them; it makes them a prerequisite.