The state of the art is inside the joint
Day one belonged to the companies putting machines on factory floors. Day two belonged to the companies building what goes inside them – actuator safety, fingertip sensing, controller architecture and the data pipelines that turn any of it into behaviour. The only panel of the event produced the sharpest exchange of the two days.
The factory is not the constraint. The model is.
Lyn Wang of Unitree, Jannes Moehlenkamp of P3, Yvan Bourqui of Johnson Electric and Muhammad Saeed of Arena 2036 were asked what stands between humanoids and mass production. Wang answered from the position of the manufacturer with the most units in the field: factory capacity is not the limit, robot intelligence is, since machines that dance convincingly in a proof of concept remain very hard to deploy in a real environment. Her estimate of more than a million units was placed five to ten years out and made conditional on better models rather than on production capability.
Bourqui put the economics plainly. A chicken-and-egg problem: volumes are too low to bring cost down, while cost is too high to generate the demand that would create volume, and the thousand-dollar unit target is still far off. His two levers were standardisation, which aggregates volume across customers, and flexible manufacturing, which avoids retooling for small specification changes. His most useful figure was that roughly eighty percent of a product’s final cost is determined during the design phase – which is why he argued for collaboration between supplier and integrator before the design is frozen rather than after. On sharing intellectual property he was cautious, describing a field too young and too competitive for voluntary openness, while pointing to telecom as the precedent where proprietary IP eventually became shared through standards everyone adopted.
Moehlenkamp argued that producibility is the neglected discipline: the shift from 3D printing to injection moulding, the move from designing for performance to designing for cost. He warned that deep vertical integration becomes a liability at scale, which cuts directly against Unitree’s stated model. His point about small and medium manufacturers is the one worth carrying home. Family-owned firms are hit hardest by labour shortage while having the least access to robotics, and P3’s answer is to build capability inside the customer’s own operations team rather than deliver a system and leave.
A single humanoid today requires six to seven people spending around three hours a day simply understanding and supervising it.
Muhammad Saeed, Arena 2036That is the number that best captures where the technology actually sits. Arena 2036 runs two collaboration formats against it: speedboat projects that take one defined problem to a deployable proof of concept in six months, and platform projects where four or five large companies work with research institutes across several challenges at once. Saeed also placed academia’s obligation precisely – carry solutions from TRL 6 to TRL 7 through 9 alongside industry, rather than stopping at publication.
Two exchanges from the floor deserve recording. A question on functional safety pointed out that verifying an actuator followed its commanded trajectory proves nothing if the robot’s brain issued the wrong command in the first place, which suggests the safety framework for humanoids needs rethinking at the system level rather than the joint level. And Saeed described an AGV deployment at Marienhospital Stuttgart that surfaced around a thousand practical problems, including elderly patients becoming distressed by autonomous movement – the kind of finding that only appears once a machine meets real people.
Half a million robots each writing into separate data centres produces no compounding intelligence. The panel called it a brake on the whole field, and noted that China is already developing national standards for video and training data collection.
Actuator safety is becoming a standard, and the standard is being written now
Ricard Picas of Novanta gave the most technically substantial talk of the day. He opened from the fact that the actuator is up to half of a robot’s hardware cost and is no longer a commodity, since it serves as both movement provider and perception device – reporting position, force, torque and acceleration. His architectural argument favoured distributed functional safety over centralised, because reusing the same sensors for control and safety removes the latency that forces a large separation distance between human and machine.
The evolution he described runs from uncontrolled stops, where safe torque off cuts power and the robot falls so the cell must be caged, through today’s controlled stops and actively damped falls, where regenerative braking is configurable per joint and the direction of a fall can be detected and managed, toward a future state where the machine does not fall at all because torque is maintained through failure and stability is measured by IMUs inside the actuators themselves. Loss of communication triggers a pre-programmed dynamic response rather than a collapse, and safe trajectory monitoring supervises complex paths rather than velocity and position alone.
Toward a future state where the machine does not fall at all – torque is maintained through failure, and stability is measured by IMUs inside the actuators themselves.
Ricard Picas, Novanta, on where actuator safety is headingThe regulatory point matters more than the product detail. A draft product standard covering dynamically stable industrial mobile robots is in development on a three-year project that began in July 2025 and releases in mid-2028, intended to be harmonised across the EU and the US and therefore legally binding. The working group is open in the sense that any country may appoint experts. Norway should have people in that room.
Fingertips, where three companies are converging
Mo Maghsoudnia of UltraSense Systems presented the most differentiated sensing approach we saw across both days: time-of-flight pulse-echo ultrasound inside the fingertip, inferring localised deformation at five hundred micron resolution, which matches human fingertip sensitivity, and producing a two-dimensional force map rather than the single scalar that camera vision and motor current give you today. The capability nothing else offers is material detection. Reflection amplitude varies with the acoustic impedance of whatever is being touched, so metal, rubber, glass and plastic are distinguishable at contact. Sub-surface lateral displacement gives shear and slip detection, and because the sensing element sits inside the housing the elastomer can wear without degrading electrical performance – which is the failure mode that ends capacitive e-skin after thirty to forty thousand touches.
Andreas Friedrich of Allegro MicroSystems approached the same body from automotive semiconductors, and his pain-point list matched what deployers reported the day before: shoulder joints overheating and forcing reduced lift capacity, torque ripple producing jerky motion, and acoustic noise mattering more than expected in shared factory space. His engineering answer to the thermal problem is architectural. Moving from 12V to 48V reduces thermal losses by a factor of sixteen, because loss scales with the square of current. His sensing portfolio spans inductive position sensing for large joints at a tenth of a degree, TMR for small joints in two-by-two-millimetre packages, and a quantum tunnelling tactile film claiming up to fifty-five times the sensitivity of strain gauges.
Michael Mross of HARTING showed moulded interconnect devices, where a laser activates an additive inside a plastic part and conductive traces are then plated onto the activated surface, which allows circuits, sensors, heating and heat dissipation to be integrated into the structural part itself. The demonstrator was a fingertip with embedded electronics, and his honest caveat was that assembly time remains high at this component density.
The difference between 95 and 99
Jan Schneider-Kraus, Head of Engineering at Robert Bosch Robotics, set the internal KPI for factory automation at a task success rate of 99-point-something percent, and was direct that ninety-five percent is already hard while 99.x is a different problem entirely. That puts the eighty-five percent Arcelik-LG reported from its rubber-part station the previous day into proportion. He considered locomotion largely solved for factory settings, with moving bases sufficient for most plant environments, and named manipulation as the real bottleneck – which is why Bosch is developing dexterous manipulation jointly with Schunk.
His data pyramid ran from robot operation data as the scarcest and most valuable, through embodied human data from wearables, then simulation, then internet video as the largest and lowest-quality corpus, with the stated goal of pushing embodied human data collection into the millions of hours. His warning about egocentric video is one we should repeat to buyers: inferring pose, force and tactility from video is a stochastic approximation that introduces systematic error, so the modalities have to be captured directly and synchronised in time, and quality has to be solved at the source rather than repaired downstream.
Whoever owns the data controls it.
Jan Schneider-Kraus, Robert Bosch Robotics – said without elaboration
Data quality as a discipline
Julian Kramer of Exxeta gave the talk closest to our own operating model. A five-level data pyramid trading cost against scale: teleoperation on a real robot capturing all joint and gripper states at the highest cost, handheld UMI grippers that lose joint states, motion capture suits that scale without a robot but lose gripper state, simulation that preserves data structure while losing real physics, and general image and language data at the base – with the constraint that some real embodiment data is always required regardless of what else is available.
A hundred high-quality episodes beat a thousand containing ten bad ones.
Julian Kramer, Exxeta – the argument the market has not internalisedIdle ratio should be trimmed or at minimum flagged, and around one percent failure episodes should be deliberately included and labelled with recovery sub-tasks, because the models learn to recover from them. His pipeline runs collection, quality control, curation into versioned and traceable datasets, fine-tuning, task-specific post-training, then deployment – precisely the sequence we built KBQS and chain-of-custody documentation to serve.
He also named what he called the data dilemma: manufacturing vision data is intellectual property, and European companies will not put it on non-EU servers. Exxeta’s answer is a sovereign on-premise platform covering collection, curation, training and fleet deployment. That is the same problem we address from the opposite direction, since a company that will not export its data can still license it under terms that keep control. The two approaches are complementary rather than competing.
Building the layer before the robots arrive
Jan-Andreas Duske of Swisslog, part of the KUKA group, argued that industrial automation fails for three recurring reasons: master data that does not hold, variety that breaks the flow, and a bottleneck that turns out to sit somewhere other than where automation was applied. The underlying absence is a common operating system. SAP, WMS and MES create islands, while Excel remains the largest ERP in the world and is neither trackable nor trainable.
The human form is the first automation that does not require the environment to be rebuilt around it. It opens the entire installed base, not only new construction.
Jan-Andreas Duske, Swisslog – the strongest case for humanoids we heard all weekWhat it does not solve is judgment. Knowing when a task is finished, or how to handle a damaged part, is unwritten shop-floor knowledge that has to be captured and formalised. His sequencing was unified semantics first, then unified actions, then unified data, all before hardware – on the grounds that integration timelines of six to twelve months now run slower than the AI evolves, so a system specified today is outdated at go-live. The live reference is an Ohio automotive plant running since July with two hundred and eighty-five robots bolting a body shell every two seconds, where the hard part was legacy systems, semantics and exception paths rather than robotics.
What it is now selling
Zhang Min of Unitree presented the breadth that has become the company’s argument: the R1 humanoid now at four thousand nine hundred dollars against roughly twelve thousand for H1 in 2023, ninety-five percent of components developed in-house, a new product generation every twelve to eighteen months, and more than a hundred robots working daily in their own factory assembling G1 units. The part relevant to us is that Unitree is now presenting itself as a data company as well – a six-thousand-square-metre proof-of-concept facility running material sorting, box handling, domestic and rehabilitation environments for collection and training, open-source whole-body datasets covering supermarket, home and office, and the claim that factories generate an average of two thousand terabytes per month, described from the stage as the critical raw material for physical AI. Their compute trajectory runs from around two hundred and fifty TOPS today to a target of two thousand by 2028.
Rokae presented from the controller side, positioning itself as a Tier 1 supplier to humanoid OEMs in the way Bosch and ZF supply automakers: a controller platform running industrial arms, cobots and humanoid arms in one programming environment, a new embodied-AI controller handling up to twenty axes with whole-body control so that AI teams can work on models rather than motion, and the observation that force control allows a task such as USB insertion to be learned in around a hundred cycles.
Alongside them, Krzysztof of Nucleus Robotics, a four-month-old Munich company, described a business model of selling labour by the hour rather than hardware: one to two months of simulation preparation before going straight onto the factory floor, and an autonomy split today of roughly fifty-five percent teleoperation, twenty-five percent human-in-the-loop prompting and twenty percent autonomous. They moved from bipedal to wheeled platforms quickly because bipedal certification is currently close to impossible, and they are targeting twenty-five robots deployed by the end of the first quarter of 2027.
Security, and who pays for care
William Dalton of VicOne made the security case: robotics is a more attractive target than IT because compromise produces physical harm rather than financial loss, frontier models have moved CVE exploitation success rates from around fifty-five percent six months ago to near total today, and a threat actor can run a frontier model on-premise for roughly eighty thousand dollars. His structural point is that manufacturers ship general-purpose firmware into deployments that each need their own security profile defined by expected behaviour – the same per-deployment logic Porsche Engineering applied to safety the day before.
The care sector session made the demand-side argument. Eighty-five percent of home care recipients in Germany are nursed by relatives, so relatives are the buyer, and the barriers are acceptance, privacy concern, missing wifi in nursing homes and above all the absence of any model for allocating ongoing cost once a pilot ends. The funding options proposed were rental rather than purchase, reimbursement tied to demonstrated outcomes, and cost split across insurers, authorities and providers.
Three companies, no commercial relationship, one conclusion
Bosch, Exxeta and Swisslog independently identified data ownership as the intellectual-property question that decides whether a deployment happens, and each has responded by building or buying a way to keep control inside the customer’s environment. That is the market forming around the problem we exist to solve, and it tells us the objection we will meet is not that data has no value but that its value is too obvious to release without terms.
Exxeta’s quality argument – a hundred clean episodes beat a thousand containing ten bad ones, and failure episodes must be labelled rather than discarded – is a direct external validation of scoring datasets before listing them. Bosch’s insistence on solving quality at the source is the same argument aimed one step upstream. Versioned and traceable curation appeared in Exxeta’s pipeline as a named stage, which is the first time we have seen provenance treated as infrastructure rather than as paperwork by anyone other than ourselves.
Kinetic Blocks is in Stuttgart for the rest of the week
Kinetic Blocks opened in beta on 1 September 2026 with twelve data vendors under MoU, covering egocentric human video, teleoperation recordings, and robot execution and action data. Every dataset is scored through KBQS before listing, and delivery is in LeRobot v3.0 format with chain of custody documentation attached.
Egocentric human video, teleoperation recordings, robot execution and action data.
Delivered in LeRobot v3.0 with chain of custody documentation attached.
NVIDIA Inception member. Raising in the fourth quarter of 2026.
If you record human or robot data on an industrial site and hold unused commercial rights, or if you are buying data and cannot establish where it came from, that is the conversation we came to Stuttgart for.
