The photo used in the cover is from: https://www.japantimes.co.jp/business/2026/04/28/companies/jal-humanoid-robot-use-airport/.
Physical AI Digest is a weekly briefing produced by Klaudia from Physical AI Company xBerry - a tech company based in Poland building tools at the intersection of Physical AI and operations.
Most Physical AI deployments this year happened inside factory walls. Controlled environments, known floor plans, defined task parameters. The factory is the right place to start: predictable, measurable, justifiable to a CFO with a payback model.
The JAL pilot is the signal that cuts through all of it. Physical AI is no longer only building capability inside controlled environments. It is testing whether that capability holds in the unstructured world.
Stats:
| Value | Description |
|---|---|
| 3 | Operational categories where Japan Airlines is testing humanoids at Haneda: baggage, gate transfer, cabin cleaning between flights |
| 99% | Figure 03 component placement accuracy across 30,000 BMW X3 vehicles - a production benchmark, not a pilot result |
| $38B | Total Physical AI market size per State of Robotics 2026, with 12 commercial humanoids now available or in pre-production |
| 15,000 | Cumulative humanoid units produced by AgiBot - the highest production volume among any humanoid manufacturer globally |
Why an Airport Changes the Argument
The Physical AI deployments receiving the most attention in 2026 have been factory-floor pilots: defined spaces, repeatable tasks, clear ROI models. Factories are the right proving ground. They are also the easiest proving ground. The floor plan does not change. The task parameters are documented. The people sharing the space know the robot is there and follow safety protocols designed around it.
An airport terminal is a different problem entirely. Haneda processes hundreds of thousands of passengers per month. Gate assignments change on short notice. Flights delay. Passengers behave unpredictably - moving against the flow, stopping mid-corridor, abandoning luggage. A cabin cleaning task that takes 18 minutes on a 20-minute turnaround leaves no room for recalculation.
Japan Airlines is testing across three task categories at Haneda: baggage loading and unloading between the terminal and aircraft hold, passenger and equipment transport between gates, and cabin cleaning and preparation between flights. Each category tests a different dimension of operational robustness. Baggage handling tests physical manipulation under time pressure. Gate transfer tests navigation in a dynamic human environment. Cabin cleaning tests sustained task completion in a semi-structured confined space.
What JAL is actually testing is not whether the robot can move luggage. It is whether humanoid operational reliability holds outside the factory. If it does, every hotel corridor, hospital wing, logistics terminal, and airport on the planet becomes a deployment target simultaneously.
A hotel corridor and a hospital wing and a warehouse aisle share more with an airport terminal than with a BMW factory floor. The tasks are different, but the environment type is the same: variable, human-dense, time-pressured, and unforgiving of interruption. Japan Airlines is not running an aviation experiment. It is running the first serious test of whether humanoid general-purpose mobility works in the unstructured real world at operational scale.
The Report That Sets the Baseline
The State of Robotics 2026 report from the Robotics Center of Silicon Valley provides the first consolidated market picture: $38 billion in total market size, 12 commercial humanoid platforms either available or in pre-production, and Vision-Language-Action (VLA) as the dominant model architecture for robot control. VLA replaces manual programming with demonstration-based learning: the robot learns a task by observing it being performed, rather than executing code written in advance. The implication is significant: the data flywheel now rewards companies with the most operational hours, not the best software architects.
Alongside the report, BMW and Figure AI released specific production data: Figure 03 achieved over 99% component placement accuracy in hard-to-access areas of the BMW X3 body across more than 30,000 vehicles. The precision benchmark comes from sections of the vehicle where ergonomic constraints make human repetitive work hardest to sustain. Figure 03 operated continuously in those spaces. The 99% figure across 30,000 units is no longer a pilot metric. It is a production record.
12 commercial humanoids simultaneously available to enterprise buyers is not a research category. 99% accuracy across 30,000 vehicles is not a demonstration result. These are production numbers, and they change the conversation every OEM is having with its robotics vendors.
The Infrastructure Week Behind the Headlines
Three signals from this week belong together. Amazon Web Services announced a Physical AI infrastructure platform on August 24 specifically designed to help enterprise customers move from laboratory pilots to production-scale deployments. AWS identified the primary scaling barrier as data and model management in real time, not hardware. Agility Robotics, Amazon's own humanoid unit, is already operating for GXO, Schaeffler, and Mercado Libre on warehouse tasks. For the existing AWS enterprise customer base, the path to Physical AI now runs through infrastructure they already pay for monthly.
Schaeffler committed to deploying 1,000 to 2,000 Neura Robotics humanoids across its global factories by 2032, with first deliveries planned between December 2026 and June 2027 at two locations: box handling in Herzogenaurach and full-scale factory testing in Schweinfurt on bearing components for electric vehicles requiring micrometer tolerances. Schaeffler participated as an investor in Neura's $1.4 billion round in June 2026 and is simultaneously its first named customer. The 2032 number with delivery dates is the first CAPEX-level commitment from a major industrial manufacturer: not a pilot budget, but a transformation schedule.
Unitree debuted on China's STAR Market at +460% on its first day of trading - among the most significant first-day performances in Chinese tech market history. AgiBot reached 15,000 cumulative units produced - the highest production volume among humanoid manufacturers globally. Optimus Gen 3 entered low-volume ramp targeting 50,000 units by end of 2026. The public markets, the CAPEX plans, and the production volumes are converging on the same conclusion: Physical AI has crossed from pilot economics into manufacturing economics.
What to Watch Next
- JAL Haneda outcome metrics: task completion rates, turnaround time impact, and whether JAL discloses the robot manufacturer - the first operational data from an airport deployment will define the benchmark for every non-factory sector
- Figure 03 expansion beyond BMW: the 30,000-vehicle production record makes Figure a credible candidate for OEM contracts outside automotive - which sector contracts next indicates where 99% placement accuracy matters most
- VLA adoption outside China: the State of Robotics 2026 report names VLA as the dominant architecture, but adoption has been led by Chinese platforms - the first Western humanoid manufacturer to commit to a VLA-based production system will be the next marker
- AWS first enterprise Physical AI customer disclosure: a named customer using the AWS deployment platform will confirm whether the infrastructure approach reduces time to production or simply shifts the integration problem elsewhere
- Schaeffler December 2026 delivery: the most important near-term signal for enterprise buyers evaluating their own procurement timelines - first confirmation that a major manufacturer's CAPEX robotics commitment is executing on schedule
FAQ
Q: Why does it matter that Japan Airlines is testing at an airport rather than a warehouse or factory?
Factories and warehouses share a defining characteristic: they are designed around the machines that operate in them. Floor layouts are planned, access paths are marked, task parameters are documented, and the humans who work alongside robots follow safety protocols that assume robot presence. An airport terminal has none of these properties. Gate assignments change on short notice. Thousands of passengers move through shared space without safety awareness training. Turnaround schedules leave no buffer for robot recalibration or human handoff. A humanoid robot that performs reliably at Haneda has proven something categorically different from one that performs reliably on a BMW production line: that the operational model holds in environments that were not designed around it. That is the precondition for deployment in hotels, hospitals, retail, and any other high-variability human environment.
Q: What is VLA and why does the State of Robotics 2026 report identify it as the new control standard?
Vision-Language-Action, or VLA, is a model architecture that trains robots to perform tasks through observation and demonstration rather than explicit code. Instead of writing a program that instructs a robot how to pick up a specific component in a specific configuration, you show the robot the task being performed and the model learns the mapping from visual and linguistic inputs to physical action. The State of Robotics 2026 report identifies VLA as dominant because it shifts the scaling advantage from software engineering headcount to data volume: the robot improves as it accumulates more operational hours, not as engineers write more code. This creates a compounding advantage for operators with large fleets, many deployment hours, and diverse task environments.
Q: Schaeffler is planning 1,000 to 2,000 robots by 2032 - is that a significant number for the industry?
In the context of where the humanoid industry stands in mid-2026, yes. AgiBot has produced 15,000 cumulative units total - the highest production volume of any humanoid manufacturer globally. A single industrial customer committing to between 1,000 and 2,000 units from one vendor represents a procurement equivalent to roughly 7 to 13 percent of the current global cumulative production leader's total output. More importantly, Schaeffler is not a technology company or a logistics operator running an AI pilot. It is a precision industrial manufacturer with over 150 years of history, an existing investor relationship with Neura, and specific factory locations and delivery dates named in the commitment. When a company that translates CAPEX decisions into five-year manufacturing plans attaches delivery dates to a humanoid purchase order, it is using the same decision vocabulary it used for every previous technology adoption in its history. That is the signal: not the number of units, but the category of decision-maker who made it.
