The new job is the real test for humanoid robots

The cheapest place to try a new job is a copy of the factory. BMW says checking a new model once meant guiding a real car body by hand through its production lines, often over several weekends, and the paint shop's dip tanks sometimes had to be emptied and cleaned. In its Virtual Factory, with digital twins of more than 30 production sites, BMW says the same check now takes three days instead of almost four weeks.

ISO 23247, the international framework for digital twins in manufacturing, defines a twin by its synchronization with the thing it represents, not by its 3D model. A one-off clearance check can do without that. Jobs run against a plant that keeps changing cannot, and a 2023 study of 29 twin architectures described in research papers found 12 that implemented synchronization and two that kept versions of the twin.

Physical AI, robot models trained on large amounts of demonstration and video data, goes after the next cost: teaching the robot the job. Google DeepMind said in June 2025 that its on-device Gemini Robotics model adapts to new tasks with as few as 50 to 100 demonstrations. Physical Intelligence wrote in April that for its pi0.7 model, seen tasks often have success rates above 90%, and unseen tasks 60 to 80%. Both are vendor measurements, on tasks the vendors chose.

The outside test found for this note did not cover either model. ArmnetBench, run in July 2026 by a company that does not sell the models it tested, trained or fine-tuned seven models on the same 50 demonstrations per task. The top scorer, Physical Intelligence's earlier pi0.5, averaged under 50%, and on its low-cost arms no model succeeded at pushing a cable into a holder. The exam is expensive too. In a box-assembly task its paper calls a factory deployment scenario, Physical Intelligence's own method used 600 autonomous trials and 360 with human interventions in each of two improvement rounds. A faster learner still has to pass the same exam: the job done on every shift without a person stepping in.

The humanoid's pitch is that the station need not change shape for the robot. The case this journal keeps returning to is Figure 02 at BMW's plant in Spartanburg, South Carolina, in what BMW calls a pilot, loading sheet-metal parts onto a welding fixture. Figure and BMW both report more than 90,000 parts moved in about 1,250 operating hours, without saying how many robots shared that total. Figure published targets of more than 99% success and zero interventions per shift, but not what it achieved.

When BMW announced a second humanoid job there in June 2026, sorting parts into sequencing trolleys, it went to Figure 03, running Figure's newer Helix 02 model. BMW says the robot will now start on the work, and Figure calls it a first demonstration. The change of robot says little on its own. Figure had begun retiring Figure 02 in November 2025, and it lists a logistics deployment among that robot's firsts. This research found no published figure for what an added task cost in time, data or interventions. What is being built in public is a place to teach. On 21 September Boston Dynamics opened a training center on the campus of Hyundai's Metaplant in Georgia, where Atlas starts with parts logistics and sequencing, with component assembly to follow by 2030.

Industrial automation has long struggled with this. In interviews MIT published in 2020 with nine organizations in Germany, France and Italy, from robot makers to large manufacturers, the difficulty of repurposing robots came up as a common theme across all of them. A luxury carmaker said that once a line's robots are dismounted, it often proves more economical to dispose of them. Yoram Koren, who by his own account proposed reconfigurable manufacturing in 1995, named the mismatch underneath: a plant runs for 12 to 25 years, while demand is forecast up to about eight years ahead.

The nearest public test of a portable process this research found comes from cell therapy, and it is a partial one. Cellares' chief executive said in January that a process qualified on its automated platform in South San Francisco "runs the same in New Jersey, Leiden, or Japan," though the Leiden and Japan sites have not opened. By Cellares' account, bringing one partner's therapy onto the platform took three years of adapting the process and generating comparability data. In late August it became public that Bristol Myers Squibb was ending an agreement worth up to USD 380 million, saying the platform could not meet the requirements to make its approved therapy, Breyanzi. Cellares says BMS ended the work before data for the FDA could be generated. Both accounts put the difficulty in the same place: carrying an approved process onto new equipment and showing it still meets its requirements.

For this site's thesis, that cuts both ways. Moduloa's bet is portable, certified production capacity, with humanoid robots as the flexible layer and the lasting value in the framework around them: standards, data, certification, training and routing. The evidence here shows only that this layer is missing, which marks a gap, not proof that it holds value. Against it, hardware abstraction, synchronized twins and fast-adapting models could make requalification a routine software check, and the companies already selling the twin, the simulator and the system of record would be better placed than any neutral framework to hold the record.

P-27 in the register, still the nearest open deadline at 30 June 2027, gets nothing again: this research, run up to 2 October, found no certified process transferred between robotic cells owned by two organizations. On 22 September Intrinsic released core parts of its robot software as open source and says arms, grippers and sensors can then be swapped without rewriting drivers. That abstracts the machine, not the qualification. The Cellares dispute was a platform conversion, not the transfer P-27 describes. P-10, validated blueprints moving between certified factories, waits on the same step.

P-12 predicts that a factory OS, running a plant as versioned configuration, becomes central to flexible manufacturing. Twins are the nearest thing to it, and two of 29 studied architectures versioned theirs. On P-02, that humanoids are first widely adopted in repetitive, dangerous and undesirable work, IDC estimates that nearly 25,000 humanoids shipped worldwide in the first half of 2026, with research and education, performance and exhibition, and government data-collection centers taking 69%, down from 83.8% across 2025. P-18, that the moat is the framework, gets nothing in its favor: a missing record shows the layer does not exist yet, not that whoever builds it can charge for it. Nothing here is scored and the register is untouched.

What would change the reading: a deployed humanoid moved to a second task with the time, data and interventions published, Figure 03 results from Spartanburg, or a qualified process accepted unchanged in a robotic cell owned by someone else.

Ford: Ford Fathom on Track, Universal EV Production System Comes to Life at LAP ·TechCrunch: Ford throws out Henry Ford's assembly line to make low-cost EVs in America ·GM Authority: Ford Escape Production Ends, One Less Rival For Chevy Equinox To Worry About ·BMW Group PressClub: BMW Group scales Virtual Factory ·NIST: Manufacturing Digital Twin Standards, G. Shao ·Ferko, Bucaioni, Pelliccione and Behnam, IEEE ICSA 2023: Standardisation in Digital Twin Architectures in Manufacturing ·Google DeepMind: Gemini Robotics On-Device brings AI to local robotic devices ·Physical Intelligence: pi0.7, a Steerable Generalist Robotic Foundation Model with Emergent Capabilities ·Physical Intelligence: pi*0.6, a VLA That Learns From Experience ·Armnet: ArmnetBench v0.1, Parallel Real-World Evaluation of Manipulation Policies on a Low-Cost Arm Farm ·Figure: F.02 Contributed to the Production of 30,000 Cars at BMW ·BMW Group PressClub: BMW Group to deploy humanoid robots in production in Germany for the first time ·BMW Group PressClub: BMW Group advances the use of Physical AI in production with Figure 03 project in Spartanburg ·Figure: F.03 Arrives at BMW ·Figure: F.02 Decommission ·Boston Dynamics: Boston Dynamics Opens Robotics Metaplant Application Center to Train Humanoid Robots for Manufacturing Tasks ·Sanneman, Fourie and Shah (MIT): The State of Industrial Robotics, Emerging Technologies, Challenges, and Key Research Directions ·Koren, Gu and Guo: Reconfigurable manufacturing systems, principles, design, and future trends ·BioProcess International: Cellares hits $600m investment milestone to scale automated smart factories ·Cellares: First patients dosed with Cabaletta Bio's rese-cel manufactured on Cellares' automated Cell Shuttle platform ·BioPharm International: As Companies Flee Ex Vivo Cell Therapy, Even Automated Manufacturing Platforms Face Validation Hurdles ·Pharma Manufacturing: Cellares defends its Cell Shuttle platform after Bristol Myers Squibb ends partnership ·Intrinsic: Introducing Intrinsic Core, an open source approach to Physical AI ·SiliconANGLE: Google's robotics unit Intrinsic open-sources its foundational infrastructure for intelligent robots ·IDC: Half-year shipments of 25,000 units, up 432.1%, as global humanoid robot commercialization heats up (in Chinese) ·NIST: Research Opportunities for Advancing Measurement Science for Manufacturing Robotics, GCR 24-054 ·Moduloa: You cannot certify a fall ·Moduloa: The two-dollar hour has no denominator ·Moduloa: The measurement layer arrived, and it points at the robot ·Moduloa: The standards landed below the work ·Moduloa: The robot was never the expensive part

Ford's dates are Ford's own statements. BMW's collision-check timing is BMW's own. The Spartanburg totals are stated by Figure and BMW, both interested parties, and no achieved success rate, intervention count or unit count has been published. Figure 03's work at the plant has no published results, and Figure's account of Figure 02's other deployments is its own. The model success rates and trial counts are vendor measurements except ArmnetBench, a preprint from a commercial firm that does not sell the models it tested, run on tabletop tasks. Its authors say its budget of 50 demonstrations per task does not measure each model's ceiling. The twin-architecture counts come from one study of architectures described in papers, not an audit of deployed factories. The Cellares account rests on Bristol Myers Squibb's statement as relayed by trade press, Cellares' rebuttal and Cellares' own releases, and the reason for the termination is contested. IDC's shipment figures are estimates. A missing published re-tasking cost and a missing cross-company transfer are negative findings from research completed on 2 October 2026, not proof that none exists. Correction, 2026-10-03: on 26 September this journal reported, via NIST, an MIT research finding that integration can cost up to five times the robot. The figure is one robot maker's estimate of four to five times, given in the MIT interviews cited above. NIST cited the MIT paper but restated the estimate as a general finding without saying who gave it, and this journal should have reported it as one company's estimate. Correction, 2026-10-03: on 23 July and 10 August this journal reported Figure 02's placement accuracy at Spartanburg as above 99 percent. That was Figure's published target per shift, not a result, and no achieved rate has been published.