DEV Community: xBerry

Amazon crossed 1 million robots this week. Then Skild AI raised $1.4 billion to make them all smarter.

xBerry — Tue, 07 Jul 2026 08:10:11 +0000

Physical AI crossed three thresholds this week: proof at true industrial scale, the foundation model race, and a first publicly traded pure-play for retail investors.

Value	Description
1M	Amazon warehouse robots in June 2026, with DeepFleet AI delivering 10% efficiency gain across the global network
$1.4B	Skild AI raise: one foundation model architecture for every robot on every hardware platform
$935M	Apptronik Series round at $5.5B valuation, tested by NASA and Mercedes-Benz
64%	Of all commercial Physical AI deployments concentrated in logistics, food service, and semiconductor manufacturing

Amazon's Million Robots and What Scale Actually Looks Like

There is a specific moment when a technology shifts from deployment story to infrastructure story. For Physical AI, that moment arrived with Amazon's confirmation that its warehouse robot fleet surpassed 1 million units in June 2026.

The number alone is remarkable. What makes it a structural signal is the layer running above it. Amazon's DeepFleet AI system, deployed across the same network, uses machine learning to coordinate routing and optimize transport across the entire fleet, delivering a 10% efficiency gain at global scale. 1 million robots plus real-time AI coordination is not a scaled-up pilot. It is a new logistics infrastructure.

No other company has deployed Physical AI at this scale with this level of centralized intelligence. Amazon is simultaneously the largest customer, the largest operator, and the most advanced real-world training environment for Physical AI systems. The operational data generated by a million coordinated robots is an asset with no equivalent in any research lab or competitor's warehouse.

Skild AI's $1.4 Billion Bet on the Foundation Model for Every Robot

The Amazon deployment answers what Physical AI looks like at scale. Skild AI is betting $1.4 billion on answering a different question: what is the foundation model layer that makes it possible for every robot to learn every task?

Skild AI closed a round of $1.4 billion, bringing total funding past $2 billion, with a mission that the robotics industry has been circling for years: a single AI architecture that operates across different robot hardware without reprogramming. The goal is to eliminate the cost of specializing AI for each new robot platform. If Skild achieves it, deploying a new physical robot becomes as straightforward as deploying a new application on an existing operating system.

Before GPT-scale models, every NLP application required its own training pipeline, its own dataset, and its own engineering team. After foundation models, the same base architecture serves translation, summarization, coding, and reasoning. Skild is attempting the same abstraction for physical action. Whoever owns the foundation model for Physical AI sets the rules for every application built on top of it.

Why the foundation model race matters for buyers: If a general-purpose robot foundation model succeeds, the cost of deploying a new robot for a new task drops from months of custom training to days of fine-tuning. Every procurement decision made today should include an assessment of which platforms will be compatible with the emerging foundation model ecosystem.

Agility Robotics Goes Public: Physical AI Reaches Retail Investors

Agility Robotics announced plans to go public via a SPAC merger with Churchill Capital Corp XI. If the transaction closes, Agility becomes the first pure-play humanoid robot company available to retail investors on public markets.

Agility's Digit robot is working commercial shifts at Amazon and Toyota Motor Manufacturing Canada under Robot-as-a-Service contracts. The CEO's statement at announcement was notably precise: not promising a robot in the home anytime soon.

Apptronik closed $935 million at a $5.5 billion valuation, tested by NASA and Mercedes-Benz. AI2 Robotics from Shenzhen raised $735 million at $3 billion with a wheeled humanoid targeting mass deployment markets.

The State of the Market: Where Physical AI Is Actually Deployed

The State of Robotics 2026 Report provides the clearest quantitative picture: a $38 billion market, 12 commercial humanoid platforms available for purchase, and logistics, food service, and semiconductor manufacturing accounting for 64% of all commercial Physical AI deployments.

Japan Airlines signing a 3-year operational contract for humanoid robots at Haneda Airport extends the deployment logic into aviation. When an airline with strict safety certification requirements signs a multi-year operational contract, it signals the technology has passed a compliance threshold, not just a performance one.

What to Watch Next

Agility SPAC closing timeline: Watch the closing date and post-listing price action as the first real market signal for what retail investors think Physical AI is worth.
Skild AI first platform integration: The foundation model thesis only proves out when a major robot manufacturer integrates Skild's architecture and reports training time reduction.
AI2 Robotics Western market entry: The wheeled humanoid model targets factory and warehouse environments at a price point that could undercut Western platforms.
Apptronik Mercedes-Benz results: A public performance report would be the first data point on how a premium humanoid performs in European automotive manufacturing standards.

FAQ

Q: What does Agility Robotics going public mean for investors who want exposure to Physical AI?

Until this transaction closes, retail investors have had no direct way to invest in humanoid robotics companies: all major players including Figure AI, NEURA Robotics, and Apptronik are private. Agility as a public company creates direct exposure to a humanoid platform with commercial revenue from real industrial deployments.

Q: What is Skild AI building and how is it different from NVIDIA Cosmos 3?

Skild AI is building a foundation model for physical action: a single AI architecture that can be deployed across different robot hardware platforms without reprogramming each platform separately. NVIDIA Cosmos 3 generates synthetic training environments to accelerate robot learning. They address different constraints: Skild attacks hardware fragmentation, Cosmos 3 attacks real-world data scarcity.

Q: Why are logistics, food service, and semiconductor manufacturing the leading deployment verticals?

These 3 sectors share the conditions that make Physical AI ROI calculable today: repetitive and physically defined tasks, high labor costs relative to robot operating costs, and environments structured enough for current robot capabilities.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

SoftBank just built the World's Biggest Robot Empire. Here's what you missed this week.

xBerry — Fri, 03 Jul 2026 09:05:45 +0000

The first week of H2 2026 did not bring incremental news. It brought a structural reset across every layer of Physical AI simultaneously: chip architecture, hardware consolidation, national policy, and enterprise proof.

Value	Description
$5.4B	SoftBank acquisition of ABB Robotics: first vertically integrated Physical AI stack from industrial to humanoid
10M	AI robots in Japan's national mandate: Physical AI becomes state infrastructure
80%	of 3,200 global leaders surveyed by Deloitte plan Physical AI deployments within 2 years
50%	Efficiency gain at Renault: 85 Exotec robots, 107,000 orders per day in German distribution center

H2 Opens With a Chip War and a $5.4 Billion Acquisition

The first structural signal of H2 2026 arrived Monday: Qualcomm introduced the Dragonwing IQ10, a processor designed specifically for humanoid robot compute. The strategic move is not just a product launch. Qualcomm is co-defining the next-generation compute architecture with Figure AI and Neura Robotics as design partners. The Dragonwing IQ10 combines strong VLA inference with low power draw, critical for robot autonomy between charges.

Until this week, NVIDIA had no credible challenger for Physical AI compute. Qualcomm changes that. Two companies now offer dedicated silicon for humanoid robots, which means buyers have architectural choices and both companies have competitive pressure to improve. The chip war for humanoids has started.

One day later, SoftBank confirmed the acquisition of ABB's robotics division for $5.4 billion. ABB Robotics is one of the world's leading industrial robot manufacturers: arms, cobots, pick-and-place systems, installed in factories across every major manufacturing economy. Combined with Boston Dynamics already in the SoftBank portfolio, the acquisition creates the first vertically integrated Physical AI stack: industrial automation hardware, humanoid platforms, and AI software under one owner.

Why consolidation matters now: A company that owns both the industrial robot installed base and the humanoid platform has a fundamentally different sales conversation. SoftBank with ABB and Boston Dynamics can walk into any ABB customer and offer a 10-year roadmap from current cobots to next-generation humanoids. No competitor can do that yet.

NVIDIA Cosmos 3: The Training Data Problem Gets a New Answer

Given Qualcomm's entry into the chip market, NVIDIA's answer this week was not a faster GPU. It was a different kind of weapon entirely.

NVIDIA announced Cosmos 3, described as the first foundation model unifying synthetic environment generation, vision reasoning, and action simulation for robots in a single stack. The core capability: Cosmos 3 generates training environments on demand, allowing robots to learn in thousands of simulated scenarios before touching a real object. The gap between concept and deployment shrinks from months to days.

This is a direct attack on the training data bottleneck that has limited Physical AI scaling. Real-world data collection is slow, expensive, and dangerous for early-stage robots. Synthetic data generated at scale removes that constraint. The company that controls synthetic world generation for robot training occupies the same strategic position that dataset providers occupied in language model development, with one critical difference: NVIDIA is not just providing the data, it is building the model stack that runs on that data.

Samsung's move to become the largest shareholder in Rainbow Robotics this week reinforces the same theme from a different angle. South Korea now has its own vertically integrated Physical AI path: Samsung manufacturing and sensors, Rainbow Robotics humanoid and cobot platforms. A third geographic vector, beyond China and the US-European axis, is building its own stack rather than licensing one.

Japan Makes Physical AI State Infrastructure

The most strategically significant announcement of the week arrived from Tokyo. Japan announced a plan for a sovereign AI model and a national target of 10 million AI robots deployed across the country. This is not a corporate roadmap. It is a government mandate.

The demographic logic is direct. Japan has one of the most aged populations in the world, a structural labor shortage across manufacturing, healthcare, and logistics, and a technological tradition in industrial robotics through Kawasaki, Fanuc, and Honda. The sovereign AI model component means Japan is not willing to run critical national infrastructure on foreign model stacks. Physical AI, in the Japanese government's framing, is the same category of strategic asset as energy infrastructure or semiconductor supply.

The parallel with China's 10,000-unit Work Mode mandate is instructive. Both governments moved from observation to mandate within the same six-month window. The difference is scale: China's mandate is 10,000 units by end of 2026; Japan's target is 10 million. As SiliconANGLE observed the same week, heavy industry is the real proving ground for this transition: structured environments, defined problems, measurable KPIs. Physical AI is no longer a market category. It is industrial policy.

The ROI Is Already Here. 80% of Companies Are Coming.

While the geopolitical and architectural stories dominated headlines, the operational evidence this week was equally significant for anyone making deployment decisions.

Renault reported the results of its February 2026 deployment of 85 Exotec Skypod robots in a German distribution center: 107,000 orders processed per day with a 50% increase in operational efficiency. The Exotec Skypod is not a humanoid. It is a vertical AI-driven storage system operating at up to 12 meters. The Renault numbers matter precisely because they are non-humanoid: they demonstrate that Physical AI delivers measurable ROI now, in standard logistics environments, without waiting for general-purpose robots.

Agility Robotics reported positive results from expanded Digit deployments in distribution centers, with reliable navigation and manipulation alongside human teams at commercial SLA. The RaaS model removes the capital expenditure barrier, turning a robot deployment into an operating cost decision, which is a fundamentally different conversation in any CFO's office.

The enterprise context came from Deloitte's survey of 3,200 global business leaders: 58% are already using Physical AI in operations. That number rises to 80% within 2 years. The barrier is no longer technological. It is organizational and decisional. Companies without a Physical AI plan in mid-2026 will be in the minority within 24 months.

What to Watch Next

Qualcomm IQ10 vs NVIDIA Isaac adoption split: Figure AI and Neura Robotics are IQ10 design partners. Which other platforms follow, and whether NVIDIA responds with dedicated low-power inference silicon, will define humanoid compute architecture through 2028.
SoftBank ABB integration timeline: The strategic value of combining ABB's industrial installed base with Boston Dynamics' humanoid platform only materializes with joint customer announcements. Watch for the first migration roadmap offer to an existing ABB customer.
Japan sovereign AI model architecture: How Japan builds its national AI model for robotics, and whether it licenses from or competes with NVIDIA, is the geopolitical AI story of H2 2026.
Cosmos 3 synthetic data adoption rate: If robot manufacturers adopt Cosmos 3 for training, NVIDIA controls the data layer of Physical AI. The next 6 months determine whether the industry converges or fragments.
Deloitte 80% by 2028 accountability: The adoption forecast creates a benchmark. At the end of 2028, the actual number will either confirm or refute the current wave of enterprise commitment.

FAQ

Q: Why does Qualcomm entering the humanoid chip market matter if NVIDIA already dominates?

A: NVIDIA's dominance in Physical AI compute has been largely uncontested because no competitor offered silicon designed specifically for robot inference requirements: real-time VLA processing, low power draw for untethered operation, and edge deployment without cloud dependency. Qualcomm's Dragonwing IQ10 addresses all three requirements and is being co-designed with Figure AI and Neura Robotics rather than sold as a generic chip. Co-design relationships create architectural dependencies that are difficult to switch, so Qualcomm is not just selling a chip but attempting to become the reference compute platform for next-generation humanoids. Competition forces NVIDIA to improve and price more aggressively, which benefits everyone building robots.

Q: What does SoftBank owning ABB Robotics and Boston Dynamics mean for industrial buyers?

A: For companies currently running ABB industrial automation, the acquisition creates a strategic question: does SoftBank use the ABB customer relationship to accelerate Boston Dynamics humanoid adoption, and if so, what does a multi-year migration roadmap look like? For buyers evaluating robot platforms now, SoftBank's vertical integration means any procurement decision involving ABB or Boston Dynamics involves the same parent company's commercial interests. It also means SoftBank has incentive to develop interoperability between the two platforms, which could create a migration path from classical industrial automation to adaptive Physical AI that no competitor currently offers.

Q: Does Deloitte's 80% adoption forecast mean most companies should be moving now?

A: The forecast describes intent, not readiness. Deloitte's data shows 80% of surveyed leaders plan Physical AI deployments within 2 years, but organizational readiness is the constraint that determines whether intent translates into successful deployment. The practical implication of the 80% figure is competitive: if that fraction of the market is actively evaluating and deploying, companies that delay lose relative position in building operational experience, training data, and process integration. The question is not whether to move, but whether to move with a readiness foundation or without one. Companies that invest in operational preparedness alongside technology evaluation will absorb deployments faster and reach Wave 2 capability sooner.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

Physical AI raised $55.8 billion in six months. The Robots are ready but your company might not be.

xBerry — Tue, 30 Jun 2026 08:29:30 +0000

H1 2026 closed with a record. NVIDIA standardized the model stack. China's Robotera raised $200M. And the sharpest post-Automate analysis asked a question nobody in the exhibit hall wanted to hear.

Value	Description
$55.8B	Raised by robotics sector in H1 2026, nearly double the full-year record from 2025
12	Commercial humanoid platforms available for purchase today, from $15,400 to $245,000
350+	Figure AI units delivered to industrial customers at one robot per hour
$200M+	Robotera Series raise: China runs its own humanoid race on its own timeline

H1 2026: The Numbers That Closed the Debate

There is a version of the H1 2026 story that is easy to tell. It goes: $55.8 billion raised, nearly double the full-year record from 2025. 12 commercial humanoid platforms available for purchase today. Figure AI delivering 350+ units to industrial customers at one robot per hour. Barclays forecasting $200 billion in market size by 2035. KraneShares confirmed the sector has officially entered its scaling phase, declaring the "race from pilot to platform" officially underway.

That story is accurate. It is also incomplete.

The harder story is the one that Tulip.co told in their post-Automate analysis, and we will get there. But the numbers deserve a moment first, because they represent something genuinely new: Physical AI in H2 2026 starts from deployment schedules, not pilot proposals. Schaeffler begins humanoid shifts in December. Toyota runs Agility's Digit on a commercial RaaS contract. Figure's BotQ ships a robot every hour. The debate about whether humanoid robots work in industrial settings is over.

NVIDIA VLA Goes Global: One Model Stack, Every Platform

The week after Automate 2026, NVIDIA announced the next layer of its Physical AI strategy: new VLA models released simultaneously with global hardware partners, each unveiling next-generation robots built on the same shared model foundation.

The new Vision-Language-Action (VLA) models bring improved spatial context understanding and longer task-planning horizons. More important than the technical specifications is the distribution pattern: hardware manufacturers across Asia, Europe, and the US all building on the same NVIDIA Isaac stack at the same time.

This is the infrastructure play that defines long-term winners. NVIDIA is not competing with robot manufacturers. It is becoming the platform they all run on. When a company's models are embedded in every robot from every manufacturer in every market, they sell infrastructure, not hardware. The same logic that made NVIDIA dominant in AI software now applies to Physical AI.

Why this matters for buyers: If you are evaluating which humanoid platform to pilot in Q3 2026, the NVIDIA Isaac compatibility of your shortlist now matters as much as the hardware specs. A robot that runs on Isaac inherits every future model improvement automatically.

China's Robotera Raises $200M: The Race Is Running on Two Tracks

While the Automate 2026 conversation focused on Figure, NEURA, and Atlas, a different signal arrived from China. Robotera closed a funding round of over $200 million, adding to a Chinese humanoid ecosystem that is running its own race on its own timeline.

The Robotera round is not an isolated data point. It is part of a pattern: Chinese humanoid companies are not copying Western platforms. They are building for a domestic market that has a government mandate (10,000 humanoids in real operations by end of 2026), local manufacturing cost advantages, and a different customer profile. Where Western platforms optimize for premium industrial applications, the Chinese ecosystem optimizes for volume and accessibility.

The implication for the global market is structural. Whoever controls training data from Wave 1 deployments gains the model improvement advantage for Wave 2. China is generating that data at state-mandated scale. The humanoid race is simultaneously a technology competition and a data accumulation race, and it is running on two tracks at once.

The Biggest Blindspot: Why Technology Readiness Is Only Half the Problem

The most important analysis of the post-Automate week did not come from a robot manufacturer or a financial analyst. It came from Tulip.co, whose "The Biggest Blindspot" report identified what the industry was systematically ignoring: operational readiness.

The argument is precise. Deploying a humanoid robot is not an IT project. It is a transformation of processes, roles, and performance metrics. A factory that buys a robot without redefining the workflows around it, retraining the workers who interact with it, and updating the KPIs that govern that production area will not fail at the technology level. It will fail at the organizational level.

This pattern has a precedent. The cloud computing adoption wave of 2011-2015 produced a familiar sequence: enterprises bought AWS capacity, then spent 18 months figuring out what to do with it. The technology was ready. The organizational absorption was not. Physical AI is moving faster than cloud, but the absorption problem is the same. Companies that invest now in operational readiness, including process redesign, workforce transition planning, and data governance for robot-generated outputs, will deploy faster in 2027 than companies that buy hardware without that preparation.

The BCG three-wave model published the same week makes this concrete. Wave 1 (now): structured task automation in predictable environments. Wave 2 (2027-2029): adaptation to semi-structured environments, which depends on training data from Wave 1 deployments. The companies that run Wave 1 pilots now are not just automating tasks. They are accumulating the data advantage that determines Wave 2 capability.

What to Watch Next

Schaeffler December 2026: First humanoid shifts in Herzogenaurach and Schweinfurt. The first large-scale test of whether BMW's 99% accuracy benchmark generalizes to a different manufacturing context.
China Work Mode November checkpoint: The MIIT progress report on the 10,000-unit deployment mandate is the first real accountability moment. Whether it lands on target will define whether the mandate accelerates or stalls.
NVIDIA Isaac partner adoption rate: With new VLA models released across global partners simultaneously, the signal to watch is how fast manufacturers outside the launch cohort integrate the stack in the next 6 months.
Operational readiness as procurement criteria: Watch whether purchasing teams start asking for operational readiness audits alongside hardware specs. If they do, Tulip.co's thesis has entered the buying process.
First public humanoid pure-play IPO: With no pure-play public humanoid company yet, watch for an IPO announcement from Figure AI, NEURA, or Agility Robotics as the next structural market signal.

FAQ

Q: What does $55.8 billion raised in H1 2026 actually mean for companies evaluating humanoid deployments?

A: The funding scale signals that the technology is past the point of existential risk: the companies building these platforms have enough capital to reach commercial maturity regardless of any single deployment outcome. For a company evaluating a pilot, this removes the "will the vendor still exist in two years?" question from the risk register. It also means the competitive pressure to move is real: competitors who pilot now accumulate Wave 1 operational data that improves their Wave 2 model performance, compounding the advantage over late movers.

Q: What is the "operational readiness" problem that Tulip.co identified, and how does a company address it?

A: Operational readiness refers to an organization's preparedness to absorb a humanoid robot deployment beyond the technical installation: workflow redesign around the robot's capabilities, workforce transition for the roles that shift, updated performance metrics that reflect robot-human collaboration rather than human-only baselines, and data governance for the operational data the robot generates. A company addresses it by running an operational readiness assessment before procurement, covering process mapping, role impact analysis, and KPI redesign. Tulip.co's core argument is that companies who buy the hardware first and figure out the organization second will underperform relative to those who prepare both tracks in parallel.

Q: Why is NVIDIA's VLA model release with global partners significant beyond the technical improvements?

A: The simultaneous release with hardware partners across Asia, Europe, and the US establishes Isaac as the shared platform rather than one option among many. In technology markets, when multiple hardware manufacturers build on the same model foundation at the same time, that foundation becomes the standard by default: the ecosystem of integrators, tools, and skills concentrates around it, making alternatives progressively harder to choose. The technical improvements in the new VLA models matter for performance, but the distribution pattern matters more for the long-term structure of the industry.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

NVIDIA set the Safety Standard. NEURA raised $1.4 billion. Barclays called $200B by 2035. Here's what you missed this week.

xBerry — Fri, 26 Jun 2026 08:47:38 +0000

Safety standards, institutional capital, and a 67-fold market forecast arrived in the same 72 hours. Here is what your organization needs to understand before Monday.

Value	Description
$1.4B	NEURA Robotics Series C from Amazon, NVIDIA, Bosch, Schaeffler, and the European Investment Bank
$55.8B	Total robotics investment in H1 2026, nearly double the full-year 2025 record
$200B	Barclays Physical AI market forecast for 2035, up from $2–3B today
67x	Projected market growth in 9 years across two deployment waves

NVIDIA Halos: Physical AI Finally Has a Safety Standard

Before this week, Physical AI had no unified safety architecture. Every robot maker built its own safety layer, and every factory deploying humanoids had to audit each system independently.

NVIDIA announced Halos for Robotics on June 22, describing it as the industry's first full-stack open safety system for Physical AI. Halos transfers the safety architecture proven in autonomous vehicles to robotics platforms, covering hardware, firmware, system software, and applications in one coherent stack. The critical design choice: Halos is open and extensible. Any robot manufacturer can integrate it into their own platform at no licensing cost.

The strategic logic is not subtle. NVIDIA is not just selling GPUs to robot manufacturers. It is establishing the safety standard that every industrial deployment will be audited against. When a COO asks whether a robot is safe to operate next to workers, the answer will increasingly reference a Halos certification.

Why this matters: Safety certification is the last mile before mass industrial deployment. NVIDIA solved the compute layer years ago. Halos completes the compliance layer, and opening it to the industry means adoption rather than fragmentation into competing proprietary standards.

The Money Became Institutional: NEURA's $1.4 Billion Changes the Capital Structure

Given that NVIDIA just established the safety standard, it is no coincidence that the same week delivered the largest full-stack robotics funding round in history.

NEURA Robotics closed a Series C of up to $1.4 billion. The investor list is the story: Amazon, NVIDIA, Qualcomm, Tether, Bosch, Schaeffler, and the European Investment Bank. This is not a venture capital round. It is a strategic alignment between technology infrastructure (NVIDIA, Amazon), industrial components (Bosch, Schaeffler), and European public capital (EBI). When the European Investment Bank writes a check for a humanoid robotics company, industrial Europe has moved from watching to committing. NEURA's valuation reached $7 billion.

That round caps a record-breaking first half of 2026. The robotics sector raised $55.8 billion in H1 alone, nearly double the full-year record from 2025. KraneShares confirmed the sector has officially entered its scaling phase, with Masayoshi Son declaring Physical AI the category that will produce the next trillion-dollar company. Capital at this scale is not speculative. It is a bet on a specific timeline.

The Warehouse Gets Its Physical AI Moment

The safety and funding announcements arrived alongside a deployment milestone that addresses the industry's most persistent gap: non-standard logistics environments.

Kawasaki Robotics and Dexterity announced an expansion of their collaboration targeting what the industry calls "long-tail warehousing": facilities handling irregular packaging, non-standard products, and variable conditions that classical automation cannot address. Dexterity's AI-driven robots handle the objects that stop conventional warehouse systems. Kawasaki brings manufacturing maturity and distribution reach. Together, they target the environments where Physical AI was still unproven at scale.

BCG published its 2026 robotics analysis the same week, identifying 3 distinct deployment waves. Wave 1, happening now: structured task automation in manufacturing, logistics, and agriculture. Wave 2, 2027-2029: adaptation to semi-structured environments. Wave 3, 2030 and beyond: general autonomy in chaotic environments. The key implication every operations leader should note: whoever controls the training data from Wave 1 deployments will dominate Wave 2 model improvement. Data from real operations at scale is the moat that cannot be replicated in a lab.

Open Models and a $200 Billion Forecast

With safety standardized, capital structured, and warehouse deployments underway, NVIDIA delivered the final piece: open foundation models that any manufacturer can use immediately.

NVIDIA published new open Isaac GR00T models enabling robots to understand natural language and execute complex, multi-step tasks using vision-language-action reasoning. The key capability: robots learn new tasks from a single demonstration, without weeks of programming. Language becomes the programming interface for industrial robots, accessible to any manufacturer worldwide at no licensing cost.

The market forecast that contextualizes all of this came from Barclays. The bank identified 2 distinct waves of humanoid deployment. Wave 1, running now to 2030: manufacturing, logistics, agriculture, construction. Wave 2, post-2030: healthcare, elder care, education, hospitality. The market is valued at $2-3 billion today. Barclays projects $200 billion by 2035. That is a 67-fold increase in 9 years.

It explains why valuations for companies like Figure AI ($39B), NEURA ($7B), and Prometheus ($41B) look rational against current revenues: investors are not pricing today's sales. They are pricing 2030-2035 market position.

What to Watch Next

Schaeffler December 2026: First humanoid robots start shifts in Herzogenaurach and Schweinfurt. Will BMW's 99% accuracy benchmark hold in a different manufacturing context?
NEURA first commercial deployment: At $7B valuation and $1.4B in fresh capital, the next milestone is a production deployment announcement. Watch for Q3 2026.
Halos adoption by non-NVIDIA partners: NVIDIA made Halos open. Whether Boston Dynamics, Agility, and Unitree adopt it or build alternatives will determine whether it becomes the industry standard or just NVIDIA's stack.
BCG Wave 1 data ownership: Toyota (Digit), BMW (Figure), Hyundai (Atlas) are accumulating training data that will improve Wave 2 models. Watch which OEMs treat this data as proprietary versus shared.
GR00T open adoption rate: Natural-language robot programming is now free to any manufacturer. How fast this becomes the default interface is the signal to watch over the next 18 months.

FAQ

Q: What is NVIDIA Halos and why does it matter for Physical AI deployment?

A: NVIDIA Halos for Robotics is an open, full-stack safety architecture for Physical AI, transferred from autonomous vehicle technology. It covers hardware, firmware, system software, and applications in a single integrated safety stack. Industrial deployment of humanoid robots requires safety certification, and before Halos every robot manufacturer built its own safety layer independently. Halos provides a shared, auditable safety standard that reduces certification time and allows factories to compare safety architectures across different robot platforms. NVIDIA making it open means adoption spreads without licensing costs, which accelerates standardization across the entire industry.

Q: What does NEURA Robotics' $1.4 billion round signal about the state of Physical AI investment?

A: The NEURA round signals that Physical AI investment has moved from venture capital to institutional capital. The investor mix includes Amazon (strategic buyer), NVIDIA (infrastructure provider), Bosch and Schaeffler (industrial manufacturing partners), and the European Investment Bank (representing European industrial policy). Each investor has a direct commercial stake in NEURA's success, not just a financial one. When the European Investment Bank participates, it signals that the EU is treating Physical AI as strategic infrastructure, comparable to how semiconductor funding has been approached in European industrial policy over the past decade.

Q: Is Barclays' $200 billion forecast for 2035 realistic?

A: The forecast is directionally plausible given the two-wave structure Barclays describes. Wave 1 (now to 2030) covers manufacturing, logistics, agriculture, and construction, which are already generating commercial deployments with measurable results. Wave 2 (post-2030) adds healthcare and elder care, which require more general-purpose capability and different regulatory approval. The $200 billion figure depends on sustained cost reductions to the $20,000-30,000 range and safety certification processes maturing faster than regulators typically move. Both are achievable but not guaranteed on the projected timeline.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

Humanoid Robots built 30,000 BMWs and cleaned Airport Terminals for $15,400. Here's why the Pilot Era is over.

xBerry — Tue, 23 Jun 2026 09:14:37 +0000

Physical AI hit 99% accuracy on BMW X3 production, JAL deployed airport robots at $15,400, and China mandated 10,000 commercial deployments by year-end. Here is what your industry missed this week.

Value	Description
99%+	Figure AI accuracy on BMW X3 assembly across 30,000+ vehicles
$15,400	Cost per JAL Unitree airport robot: baggage, cargo, cabin cleaning
10,000	China Work Mode mandate: humanoids in real operations by end of 2026
$55.8B	Raised by robotics companies in 2026 alone, nearly double the 2025 record

The Number That Changes the Conversation

The number that changes everything is not $1 trillion in projected market value, or the $55.8 billion raised by robotics companies in 2026 alone. It is 99.

That is the accuracy rate - above 99% - at which Figure AI humanoid robots participated in assembling over 30,000 BMW X3 vehicles. The benchmark exceeds standard requirements for human operators on the same task. A CFO can read that number. A COO can sign a purchase order based on it.

This week, spanning June 19-24, 2026, during Automate 2026 in Chicago, Physical AI crossed from a category of promising pilots into auditable industrial metrics. Here is what happened and why it matters to your organization now.

The Production Proof You Cannot Argue With

Three deployment milestones arrived this week that together redefine what "industrially ready" means for humanoid robots.

BMW + Figure AI: 99% accuracy across 30,000 vehicles. Figure AI's humanoids participated in assembly of more than 30,000 BMW X3 units at accuracy rates exceeding 99% for component placement. This is not a demonstration. It is a completed quality audit with data that BMW's production teams measure against human operator benchmarks.

Figure BotQ: one robot per hour, 350+ units delivered. Figure AI's BotQ factory is producing Figure 03 at a rate of one unit per hour, a 24x throughput increase in under 120 days. Over 350 units have reached industrial customers. At current rate, that translates to capacity for 8,760 robots per year from a single production line.

Atlas and Digit start commercial shifts. Boston Dynamics began commercial shipments of electric Atlas: 56 degrees of freedom, 50 kg lift capacity, full 360-degree torso rotation, autonomous battery swapping. All 2026 units committed to Hyundai's Robotics Metaplant Application Center and Google DeepMind. Agility Robotics' Digit is working commercial shifts at Toyota Motor Manufacturing Canada under a Robots-as-a-Service agreement. Toyota pays per use, not per unit.

The $15,400 Signal That Expands Who Can Deploy

Japan Airlines, working with GMO AI & Robotics, deployed Unitree-based humanoid robots in airport operations at approximately $15,400 per unit. Tasks: baggage loading, container transport between vehicles, cabin cleaning. That price is three times cheaper than comparable western-market platforms, and it is operating in an environment that is dynamic, variable, and safety-critical.

China made the price calculus more urgent with a state-level mandate. The Ministry of Industry and Information Technology (MIIT) launched its Work Mode program requiring 10,000 commercially deployed humanoid robots in real operations by end of 2026. Local governments must submit implementation plans by end of June. Progress report due in November. China is not asking whether humanoids are ready. It is assigning quotas.

The State of Robotics 2026 report confirms the bifurcation: 12 commercial humanoid platforms now available, ranging from $15,000 for torso systems to $245,000 for full bipeds. The question is which ecosystem dominates global logistics first.

What the Capital Flow Is Telling You

Masayoshi Son said it plainly on CNBC this week: Physical AI is where the next trillion-dollar company will be built. The robotics sector has raised $55.8 billion in 2026 alone, nearly double the 2025 record.

Jeff Bezos's Prometheus raised $12 billion at a $41 billion valuation from JPMorgan Chase, Goldman Sachs, and BlackRock. Prometheus is not building a robot. It is building what Bezos calls an "artificial general engineer": software capable of automating the design and production of complex physical systems. $18.2 billion in total funding and no public product demo yet.

Tesla converted its Model S and X production line at Fremont to manufacture Optimus Gen 3, targeting 50,000-100,000 units in 2026 with stated capacity of 1 million per year. Over 1,000 Optimus units are currently learning inside Tesla's own factories on proprietary operational data no competitor can replicate.

The Capgemini Physical AI Report adds context: 79% of organizations are already engaged with Physical AI, but only 30% believe general-purpose humanoids will be production-ready within 3-5 years. The median executive horizon is 7 years. Early movers build an insurmountable data advantage in that gap.

What to Watch Next

Schaeffler, December 2026: First humanoid robots start operational shifts in Herzogenaurach and Schweinfurt. Will confirm or challenge BMW's 99% benchmark.
China Work Mode accountability: November 2026 progress report. Miss it and the mandate is noise. Hit it and the global competitive dynamic shifts.
RaaS pricing standards: Agility at Toyota is the first major Robots-as-a-Service contract for humanoids. If terms become public, they set pricing expectations industry-wide.
Figure BotQ second line: A second production line announcement ends the supply constraint argument.
Prometheus first demo: $18.2 billion raised, no product reveal. The most-watched industrial AI event in the next 12 months.

FAQ

Q: Are humanoid robots actually production-ready in 2026?

A: Yes, for well-defined, repetitive industrial tasks. BMW's 99% accuracy result and Figure's 350+ delivered units demonstrate that narrow applications (component placement, baggage handling, cabin cleaning) are commercially viable today. General-purpose humanoid work across unstructured environments remains 3-7 years away according to Capgemini's survey of industry executives.

Q: What does a humanoid robot actually cost to deploy in 2026?

A: The range is $15,400 (Unitree-based systems, as deployed by JAL) to $90,000-$100,000 for premium platforms like Boston Dynamics Atlas. RaaS contracts from Agility Robotics offer subscription-based deployment with no upfront capex. Industry consensus for mass-market viability is $20,000-$30,000, projected for 2028-2030.

Q: How does China's 10,000-unit mandate affect Western manufacturers?

A: It creates a forced deployment cycle generating real-world operational data at a scale no other market produces in 2026. Chinese manufacturers accumulate deployment experience and training data faster, accelerating model improvement and driving down unit costs. For western OEMs, the window to build comparable data advantages is now.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

A prosthetic hand is now teaching an industrial robot & PepsiCo signed for autonomous freight. Here's what you missed this week.

xBerry — Fri, 19 Jun 2026 09:39:06 +0000

PSYONIC's prosthetic touch data is now training ABB robots. Gatik signed the first Fortune 50 commercial autonomous freight contract with PepsiCo. Burro drove Physical AI onto the construction site. Experts set $20k as the humanoid price target. And someone just called Edge AI the Windows of robotics.

This week, Physical AI crossed three invisible lines at once. A company that makes prosthetic hands figured out that the touch data from amputees is exactly what industrial robots need to learn how to grip. A Fortune 50 company signed not a pilot but a commercial contract for autonomous freight. A 44-horsepower robot drove off the warehouse floor and onto the construction site. And two separate conversations about software and pricing suggest that the next wave of robotics adoption will be driven by access, not capability.

Here is what happened, and why it matters beyond the headlines.

Value	Description
Fortune 50	PepsiCo becomes first to sign a commercial contract for autonomous freight with Gatik
$20k	Target price point for humanoid robots, Robotics Summit consensus: achievable by 2028–2030
1M hours	Burro's field experience backing the Grande 44 autonomous outdoor platform
100+	Pressure sensors per fingertip in PSYONIC's Ability Hand, now training ABB GoFa

A Prosthetic Hand Is Now Teaching an Industrial Robot How to Grip

The standard approach to teaching a robot how to handle objects has been simulation, teleoperation, or labor-intensive physical demonstrations. PSYONIC and ABB just introduced a different source of data: the hands of people who have already learned to feel again.

PSYONIC's Ability Hand is a prosthetic with more than 100 pressure sensors per fingertip. The company has been collecting kinesthetic data from users with upper-limb amputations. That data, which captures how a human hand adjusts grip pressure, contact area, and force across thousands of everyday tasks, is now being fed as training data into ABB GoFa robot arm models.

The implication is not obvious until you think about it for a moment. Prosthetic hand users have been solving exactly the problem that robot engineers have been trying to solve: how to grip objects of variable shape, weight, and texture using feedback from pressure sensors. They have been solving it in the real world, for years, across diverse populations. That dataset has no equivalent in any robotics lab.

This is a genuinely new approach to the data collection problem in manipulation. Instead of running robots to generate training data, you collect from humans whose daily lives already generate the signal you need. The ethics, the incentive structures, and the consent frameworks all need to be built carefully. But the technical direction is clear and it points somewhere important.

Fortune 50 Just Signed Its First Commercial Autonomous Freight Contract

There is a meaningful difference between a pilot program and a commercial contract. A pilot is a test. A contract is an operational commitment with financial stakes, SLAs, and consequences for non-performance. Gatik and PepsiCo just crossed that line.

Gatik's autonomous trucks will operate across PepsiCo's North American regional transport network, connecting warehouses to distribution points in a middle-mile model. No safety driver. No remote operator on standby. Commercial terms. PepsiCo is the first Fortune 50 company to sign at this level for autonomous freight, which means this is now a procurement decision made by a global supply chain organization with thousands of operational variables to manage.

The middle-mile use case is strategically important. It is a fixed route, predictable environment, and high-frequency run, which makes it the easiest category of autonomous freight to operate reliably. The fact that a Fortune 50 is comfortable putting commercial obligations behind it signals that the reliability question has been answered to the satisfaction of a legal and operations team, not just a technology team. That is a different bar.

For the industry, the signal is that autonomous freight is no longer waiting for regulatory clarity or technology maturity. It is already inside corporate supply chain planning cycles.

Physical AI Is Leaving the Warehouse

Most of the physical AI deployment conversation has been set inside four walls: warehouses, fulfillment centers, factories. This week, two events pushed the boundary outward.

Burro introduced the Grande 44, a 44-horsepower autonomous tractor built for outdoor heavy industry: construction sites, ports, agricultural operations, and facility grounds management. Behind it is more than one million hours of real-world field experience from previous Burro platforms. The Grande 44 does not need GPS precision or controlled surfaces. It navigates the kind of environments that traditional warehouse robots cannot handle.

In the same week, Einride, the Swedish operator of autonomous electric freight trucks running for Fortune 500 clients in the US and Europe, went public via SPAC. The IPO sends a specific signal: institutional investors see a path to profitability in autonomous logistics that goes beyond the humanoid robot narrative. The capital is following the deployments, not the demos.

Together, Burro and Einride represent the geographic and category expansion of physical AI. The technology is not contained to a single environment type or a single vehicle form factor. It is filling the operational gaps*wherever human labor is expensive, dangerous, or in short supply.

The $20,000 Humanoid and the Windows Moment

Two separate conversations from this week point at the same underlying dynamic: the next phase of robotics adoption will be driven by access, not capability.

At Robotics Summit 2026, a panel of humanoid robot designers converged on $20,000 as the price point at which ROI becomes accessible for a mid-sized factory. Current humanoids range from $25,000 to $90,000 depending on the manufacturer and configuration. The panel's consensus on when $20k is achievable: 2028 to 2030, contingent on breakthroughs in actuator and battery manufacturing. The framing of the conversation has shifted. The question is no longer whether price will fall. It is when, and which manufacturers will hit the threshold first.

In parallel, Jason Seawall made the case that Edge AI middleware is to robots what Windows was to personal computers. Before Windows, operating a PC required an engineer. After Windows, anyone could use one. Before Edge AI middleware, deploying a robot required a systems integrator and a programmer. After it, a factory floor operator can configure and run a robot without writing code. The software layer is what converts a technically capable system into something a normal business can actually buy and operate.

These two signals together describe the same future: robots that cost less and require less technical expertise to deploy. That combination is what drives mass adoption in every hardware category. It is starting to happen in Physical AI.

What to Watch Next

PSYONIC and ABB data partnership terms: whether the prosthetic-to-robot data model becomes a licensed framework that other manipulation companies can access, and what the consent and compensation structure looks like for the users generating the data
Gatik expansion beyond PepsiCo: which other Fortune 500 supply chain organizations announce commercial autonomous freight contracts in the next six months, and whether any involve last-mile rather than middle-mile routes
Einride post-SPAC performance: whether public market investors sustain confidence in autonomous freight as an investment category, and how Einride's revenue multiple compares to humanoid robotics valuations
Actuator and battery cost curves: the Robotics Summit $20k target depends on manufacturing breakthroughs that have not happened yet - the companies that crack actuator cost first will set the commercial timeline for the entire industry

FAQ: Access, Cost, and the Next Phase of Physical AI

Q: Why does the PSYONIC and ABB partnership represent a new approach rather than just another data source?

A: Most robot training data is generated by robots, which means it inherits the limitations of current robot hardware: limited sensor resolution, constrained environments, and short collection windows. Prosthetic hand users generate manipulation data continuously in the real world across years of use and across highly varied task scenarios. The density and diversity of that signal is qualitatively different from what a lab can produce. The partnership also inverts the usual direction: instead of technology being built for able-bodied users and adapted for people with disabilities, the data from people with disabilities is improving technology for everyone. That is a meaningful inversion worth tracking.

Q: What makes the Gatik and PepsiCo deal different from previous autonomous freight announcements?

A: Most autonomous freight announcements are pilots, which means the operator retains control over scope, can terminate without financial consequence, and carries no SLA obligations. A commercial contract changes all three variables. PepsiCo's procurement and legal teams approved operational commitments based on Gatik's reliability record. That approval process is more demanding than a pilot review because it involves finance, risk, and operations stakeholders who are not interested in the technology story. When a Fortune 50 legal team signs off, it means the system has passed a real-world reliability threshold, not a lab benchmark.

Q: Is the Edge AI equals Windows analogy accurate, or is it overstated?

A: It is directionally correct but the timeline is more uncertain than the analogy suggests. Windows succeeded because PC hardware was already standardized enough that a single software layer could abstract the complexity. Robot hardware is still highly fragmented: different actuators, sensors, compute platforms, and kinematics require different integration work. Edge AI middleware can reduce that burden substantially but cannot eliminate it entirely yet. The analogy captures the direction correctly: software abstraction layers are what convert technical capability into deployable products. The question is how long it takes for robot hardware to standardize enough for the abstraction to become clean. That is a five-to-ten year process, not a two-year one.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

A Robot worked a 200-hour shift. China made 10,000 Humanoid Deployments mandatory. Three Robotics Companies filed IPO the same week.

xBerry — Tue, 16 Jun 2026 08:34:57 +0000

Figure AI's Helix-02 ran 200 hours without a single human intervention. China made 10,000 humanoid deployments mandatory by year-end. Three Chinese robotics companies filed for IPO in the same week. The experiment phase is over.

Value	Description
200h	Figure Helix-02 continuous autonomous operation
149,000+	Packages sorted, zero human interventions
10,000	Humanoids China mandates in real work by end 2026
73 days	Unitree IPO approval, STAR Market record

The Experiment Is Over. Here Is What Replaced It.

Every new technology has an experiment phase and a deployment phase. The experiment phase is characterized by pilots, proof-of-concepts, and optimistic press releases. The deployment phase is characterized by mandatory deadlines, public market listings, and robots working 200-hour shifts without anyone watching.

Physical AI crossed that line this week.

What 200 Hours of Continuous Robot Work Actually Means

The question every operations director has been asking for two years is not "can a robot do this task?" The question is: "Can it do it on Tuesday, and again on Wednesday, and again on Thursday, through a full shift, without someone standing next to it?"

Figure AI answered that question with Helix-02. Three Figure 03 robots, named Bob, Jim, and Rose by livestream viewers, ran for over 200 continuous hours sorting packages. The result: more than 149,000 packages processed, zero human interventions, zero reported failures. The system used onboard cameras, AI reasoning, barcode detection, and pick-and-place to a conveyor belt.

CEO Brett Adcock's statement was precise: a full 8-hour shift at human-level performance, fully autonomously. That framing matters. "Human-level" is not a benchmark metric here. It is a commercial threshold. A robot that matches human throughput on a repeatable task, without supervision, makes the ROI calculation for a warehouse operator straightforward.

The 200-hour livestream was not a marketing stunt. It was a durability test conducted in public. Every hour that passed without intervention was evidence the system does not degrade over time. That is the data COOs need before signing a deployment contract.

The IPO Wave Is the Market Saying It Believes

Venture capital moves early and bets on potential. Public markets move later and bet on evidence. The fact that three Chinese humanoid robotics companies filed for IPO in the same week is a signal that the evidence has arrived.

EngineAI filed a confidential application for a Hong Kong listing at a valuation above 10 billion CNY. One of its facilities produces a humanoid robot every 15 minutes. Unitree received STAR Market approval just 73 days after filing, a record pace that reflects both regulator confidence and the company's financials: more than 5,500 humanoids sold in 2025, revenue of 1.7 billion CNY. Linkerbot, which focuses on robotic hands, is targeting a $6 billion valuation in its own listing.

Three IPOs in one week is not a coincidence. It is a coordinated signal from the Chinese robotics ecosystem that the companies building humanoid robots believe their revenue is real enough to justify public scrutiny. When retail investors can buy shares in a humanoid robot manufacturer, the pressure on that company to scale and hit profitability becomes permanent. That pressure accelerates the entire industry.

For Western companies watching from the sidelines, the timing is notable. Unitree's STAR Market approval came 73 days after filing. Most Western IPO processes take 12 to 18 months. The speed differential is itself a competitive signal.

China Just Made It Mandatory

While Figure AI was running its livestream and Chinese companies were filing IPO paperwork, China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission quietly announced something more consequential than either.

The "Work Mode" program sets a hard national target: 10,000 humanoid robots in real commercial deployment by the end of 2026. Not in pilots. Not in controlled environments. In representative real-world scenarios across factories, logistics, retail, healthcare, equipment inspection, and emergency rescue. Local governments must submit implementation plans by the end of June and progress reports by the end of November.

This is the first government-issued deployment mandate of this scale anywhere in the world. The framing shift is significant: China is not asking whether humanoid robots are ready. It is treating readiness as assumed and issuing a deadline. The language changed from "pilot" to "obligation."

For companies operating supply chains in or with China, this mandate has direct implications. If 10,000 humanoids are verified and deployed across Chinese factories and logistics networks by December 2026, the operational data generated will accelerate Chinese robotics models faster than any lab benchmark program could. Data from real deployments, at scale, is the input that improves the next generation of models.

The Consumer Angle Nobody Expected

Not every signal this week was about industrial scale. Faraday Future announced the launch of its EAI Robotics Education Ecosystem in Los Angeles, targeting two segments simultaneously: educational institutions (B2B) and family consumers (B2C).

The analogy Faraday Future is drawing is the school computer: PCs entered homes because children encountered them first in classrooms. The bet is that robotics education for children today creates a generation of adult consumers who are comfortable buying and living with robots. It is a long game, but it is the correct long game. Every mature consumer technology followed a similar adoption path.

Whether Faraday Future specifically has the resources to execute this strategy is an open question. The concept, however, is sound, and it will not be the last company to try it.

What to Watch Next

Figure AI deployment contracts: which logistics or e-commerce operator announces production use of Helix-02 first, and at what scale
China Work Mode progress reports: local government implementation plans due end of June - the specifics will reveal which cities and industries are moving fastest
EngineAI and Unitree IPO pricing: the valuations set in public markets will become the benchmark that every private humanoid robotics company is measured against
Automate 2026 Humanoid Robot Forum, June 22-25 in Chicago: the first major Western industry event after China's mandate announcement - expect direct comparisons
Whether any Western government follows China with a formal deployment target or procurement mandate before year-end

FAQ: From Pilots to Deployment

Q: Does a 200-hour livestream actually prove anything for industrial deployment?

A: More than most lab benchmarks do. The key variable in industrial deployment is not peak performance but consistency over time. A robot that achieves 98% accuracy in a 10-minute test and then drifts to 70% after six hours of operation is not deployable. Figure AI ran its system for over 200 continuous hours in public, where any failure would have been visible to thousands of viewers. The absence of reported failures during that period is meaningful evidence of system stability, not just capability.

Q: What does China's 10,000-humanoid mandate mean for non-Chinese companies?

A: Two things. First, if Chinese manufacturers hit the target, they will generate an enormous amount of real-world deployment data by early 2027, which feeds directly into the next generation of Chinese robotics models. Second, any company with manufacturing or logistics operations in China will encounter humanoid robots as part of their supplier or partner ecosystem within 18 months. This is no longer a future scenario to plan for. It is a near-term operational reality to prepare for.

Q: Why are three Chinese robotics IPOs happening at the same time?

A: Timing an IPO requires sufficient revenue, a compelling growth narrative, and favorable market conditions. All three appear to have converged simultaneously for EngineAI, Unitree, and Linkerbot. The broader context is China's government-backed push to dominate humanoid robotics, which has created both the capital environment and the commercial demand signal that public market investors need. The 73-day approval for Unitree suggests regulators are actively facilitating this wave, not just permitting it.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

Physical AI just got its platform layer. Nvidia is the only candidate. Here's what you missed this week.

xBerry — Fri, 12 Jun 2026 07:22:21 +0000

NEURA closed a $1.4B record round, robots grew hands that can feel, and someone is racing to own the Physical AI ecosystem.

Value	Description
$1.4BN	NEURA Series C record
$55.8B	Raised in robotics 2026
75 DoF	Sharpa Wave + Unitree
20x	Less real-robot data needed

The Week Physical AI Got a Sense of Touch, a Record Check, and a Platform War

Three days. Three storylines that change different parts of the same industry.

A German humanoid robotics company closed the largest full-stack robotics funding round in history. A startup shipped robot hands with over a thousand touch sensors per fingertip. And the question that will define Physical AI for the next decade got named out loud: who controls the body, the brain, and the ecosystem?

Robots Are Finally Learning to Feel

For three years, the dominant narrative in Physical AI has been about vision: give a robot better cameras, better vision-language models, and it will handle the physical world. The problem is that many real-world tasks cannot be solved by sight alone.

Loose cables, deformable packaging, components that shift when touched: these are the objects that stop factory robots cold. A camera sees the object. A robot without tactile feedback cannot know what its grip actually feels like.

On June 9, Sharpa announced the integration of its Wave gloves into the Unitree H2 Plus reference design on NVIDIA Isaac GR00T. The numbers: 22 degrees of freedom per hand, 75 DoF total for the full body, and more than 1,000 touch sensors per fingertip. The entire stack runs on Jetson AGX Thor, using Isaac Teleop for data collection and Isaac Lab for training.

This is not a lab prototype. It is a reference design, meaning hardware and software partners can build on it directly. The combination of GR00T's manipulation intelligence with tactile feedback closes the gap that has limited dexterous robotics for the past decade. Robots can now feel what they are holding. That sentence has not been true before now.

The Money Has Found Its Thesis

The investment thesis for Physical AI used to be speculative. This week it became structural.

NEURA Robotics closed a $1.4 billion Series C, the largest full-stack robotics round in history, at a $7 billion valuation. The investor list reads like a strategic playbook: Tether (lead), Amazon, Nvidia, Qualcomm, Bosch, Schaeffler, and the European Investment Bank. This is not venture capital chasing hype. This is industrial capital locking in supply chain relationships before the market consolidates.

Separately, Standard Bots raised $200 million at a $1 billion valuation. Their pitch: robots that learn by watching demonstrations, no coding required, 20 to 30 percent cheaper than legacy industrial players. Customers include Lockheed Martin, Amazon, and NASA. The company is advising the White House on a National Robotics Strategy.

The macro picture: $55.8 billion was raised by robotics companies in 2026, nearly double the 2025 figure. COMPUTEX 2026 opened its first-ever robotics zone. Taiwan's suppliers are pivoting from humanoid hardware to Physical AI compute platforms and edge AI. The capital is not chasing pilots anymore. It is building infrastructure.

Who Will Own the Physical AI Ecosystem

The most important question this week did not come with a press release.

Digitimes reported a debate emerging in China after Unitree launched the H2 Plus with Nvidia AI inside: who controls the body, the brain, and the training data? The comparison being made is Wintel. In the PC era, Intel owned the processor and Microsoft owned the operating system. Hardware makers built on top of both. Value accrued to the platform, not the box.

Nvidia is actively auditioning for both roles in Physical AI. Isaac GR00T provides the foundation model. Isaac Sim and Isaac Lab handle training. Cosmos generates synthetic data. OSMO orchestrates workloads. Every hardware maker that integrates these tools becomes dependent on Nvidia's stack, pricing, and roadmap.

This is exactly why Nebius and Nvidia launched a Physical AI Living Lab for European robotics startups, with the first cohort starting in September 2026. The goal is to pull the next wave of founders into the Nvidia ecosystem before competitors can establish alternatives. The company that wins the platform layer of Physical AI will collect rent from every robot sold, regardless of who builds the hardware.

The Tools Getting Cheaper While the Stakes Get Higher

Not every signal this week was about capital and control.

On June 11, X Square Robot published XRZero-G0: an open-source wearable framework that lets researchers collect robot training data without using a physical robot. The result: ten recordings with a VR headset and hand controllers plus one recording on The actual robot equals the performance of eleven robot-only recordings. The G0-Dataset contains 2,000 hours of multimodal data on Hugging Face, free to use. Code is on GitHub, paper on arXiv.

The World Economic Forum named Hello Robot a Technology Pioneer 2026 for building Stretch, a robot that helps people with spinal cord injuries perform daily tasks. CEO Aaron Edsinger's framing: the missing frame in Physical AI is the person the robot actually serves. Hello Robot measures success in total user independence, not factory throughput.

In a week dominated by billion-dollar rounds and platform debates, these two signals are a reminder that scaling and accessibility are separate vectors. Both are necessary for Physical AI to be something more than a capital-intensive industrial story.

What to Watch Next

NEURA's Neuraverse platform and NEURA Gyms: first deployment timeline and whether the decentralized AI architecture holds under production conditions
Nvidia ecosystem consolidation: which hardware partners publicly commit to full Isaac stack integration, and which hedge by supporting alternatives
XRZero-G0 adoption: whether the 20x data reduction claim holds across task categories outside the paper's benchmarks
Automate 2026 Humanoid Robot Forum, June 22-25 in Chicago, with Boston Dynamics, NEURA Robotics, NVIDIA, and Toyota Research Institute on one stage
Whether the Unitree-Nvidia "Wintel" dynamic surfaces as a formal partnership announcement or a competitive split over data and ecosystem control

FAQ: Physical AI's Platform War and What It Means

Q: What makes NEURA Robotics different from other humanoid robotics companies?

A: NEURA is building a full-stack platform: hardware, software, training infrastructure (NEURA Gyms), and a decentralized AI architecture called Neuraverse. Most competitors focus on hardware or models in isolation. The investor mix, including Bosch, Schaeffler, and the European Investment Bank alongside Nvidia and Amazon, signals that the company is being positioned as industrial infrastructure, not a consumer product. The $1 billion order book they reported alongside the raise confirms there is real demand behind the valuation.

Q: What does the "Wintel of robotics" mean for companies buying robots?

A: If Nvidia becomes the dominant platform for both training and inference in humanoid robotics, companies that buy robots built on Isaac GR00T become dependent on Nvidia's pricing and roadmap, regardless of which hardware brand they chose. For procurement and strategy teams, the vendor evaluation should include the AI stack behind the robot, not just the hardware specs. Choosing a robot in 2026 is also choosing an AI platform relationship for the next decade.

Q: Why does tactile sensing matter for Physical AI deployments?

A: Current robots rely primarily on vision. Many industrial and household tasks require force feedback: knowing whether an object is slipping, how hard to grip a fragile part, or how to handle deformable materials like cables or soft packaging. Sharpa Wave's 1,000-plus touch sensors per fingertip on the Unitree H2 Plus platform means a robot can feel the difference between gripping a circuit board and crushing it. This enables a class of tasks that camera-only robots cannot perform reliably, which covers a large share of the remaining automation gap in manufacturing and logistics.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

Physical AI has Scaling Laws now. The Race just became something else.

xBerry — Tue, 09 Jun 2026 07:31:55 +0000

NVIDIA discovered the first scaling law for robot dexterity this week. Paired with Apache 2.0 licensing, BYD's 20,000-unit push, and a $400M foundation model raise, physical AI just crossed a threshold.

Stat	Description
2x	Manipulation success rate doubles going from 1,000 to 20,000 hours of training data (GR00T N1.7)
$400M	Raised by Generalist AI on June 4 at $2B valuation, backed by NVIDIA and Bezos Expeditions
20,000	Humanoid robots BYD plans to deploy in its own factories in 2026
Apache 2.0	License for GR00T N1.7 — fully open for commercial use, no restrictions

The Week Physical AI Proved It Obeys the Same Rules as LLMs

In machine learning, a scaling law means one thing: more data and compute produce predictably better results. It is the reason GPT-2 became GPT-4 in three years. It is the reason every major AI lab now races to build larger datasets rather than better architectures.

This week, NVIDIA published the first scaling law for robot dexterity. The finding came with GR00T N1.7, released June 9 with a full Apache 2.0 license: going from 1,000 to 20,000 hours of real-world video training data doubles manipulation success rates. The model is 3 billion parameters, trained on the EgoScale dataset of 20,854 hours of egocentric video, and it does not require thousands of hours of costly teleoperation.

That one result changes the trajectory of the entire field. Physical AI no longer has to hope that more data helps. Now it knows by how much.

Open-Source Foundation Models Are Now a Reality for Robotics

Two models released this week signal a structural shift: GR00T N1.7 under Apache 2.0 and SmolVLA from Hugging Face's LeRobot team, which runs on a single consumer GPU at 450 million parameters while matching OpenVLA on standard benchmarks.

For context: the closed-source era of robot foundation models looked a lot like the closed-source era of LLMs before 2023. A handful of well-funded labs held the best models behind proprietary APIs and expensive licenses. The shift to open weights for language AI created an explosion of specialized fine-tunes, downstream products, and deployment tooling within 18 months.

The same dynamic is now beginning for physical AI. A factory engineer with a single GPU and a GitHub account can now run a manipulation model that beats proprietary baselines from a year ago. That is not a minor update. That is a platform shift.

BCG's 5-level Physical AI maturity framework, published this week, puts Amazon Vulcan at Level 4: autonomously handling 75 percent of more than one million unique product SKUs, including items it has never seen before. The framework gives operations and strategy teams the vocabulary to position their own deployments and write a credible business case for the board.

The Capital Is No Longer Speculative

Masayoshi Son told CNBC this week that physical AI is where he sees the next trillion-dollar company. That is the kind of quote that gets repeated in investor decks. What matters more is the capital already committed.

Generalist AI closed a $400 million round on June 4 at a $2 billion valuation, led by Radical Ventures with participation from NVIDIA and Bezos Expeditions. The founding team includes Pete Florence and Andy Zeng from DeepMind and Andrew Barry from Boston Dynamics. Their latest model, GEN-1, reports 99 percent reliability across diverse dexterous tasks at three times the speed of the previous benchmark. The dataset behind it: over 500,000 hours of real-world robotic activity collected via hand-mimicking grippers seeded globally.

Then there is BYD. The world's second-largest EV manufacturer confirmed on June 4 that it is developing humanoid robots under the codename Yao-Shun-Yu, a project running since 2022. 150 prototypes are already being tested inside BYD's own factories. The company plans to deploy 20,000 units internally in 2026, with a new industrial park in Xi'an targeting 50,000 units annually. Future consumer sales would go through BYD's existing dealer network. Executive vice president Stella Li put it plainly: "Automotive software is complex, and porting it into robots is very easy for us."

When the world's most efficient battery manufacturer decides to sell robots through its car dealerships, the distribution problem for humanoids is no longer theoretical.

Real-Robot Benchmarks Are Finally Replacing Simulation

The CVPR 2026 Embodied AI Workshop ran June 3-7 in Denver. This year's ManipArena competition was the first in the field scored entirely on physical robots, not simulators, across 20 distinct manipulation tasks. Three challenges ran in parallel: ARNOLD for language-grounded manipulation, ManiSkill-ViTac for bimanual vision-tactile fusion, and ManipArena for desktop and mobile manipulation.

This is a bigger deal than it looks. Simulation-to-reality transfer has been the field's unsolved credibility problem for years. Teams could rank first in a simulator and fail basic tasks on a real robot. The leaderboards from Denver now reflect actual physical dexterity. The capital will follow those rankings.

What to Watch Next

GR00T N1.7 early access: which deployment partner announces production use first, and whether independent benchmarks confirm the dexterity scaling claim
BYD Xi'an humanoid park: construction timeline and whether the 50,000 units/year capacity target holds
RoboStrategy investor presentation, June 10, covering its portfolio of Figure AI, Apptronik, and Standard Bots
Automate 2026 Humanoid Robot Forum, June 22-25 in Chicago, with Boston Dynamics, NEURA Robotics, NVIDIA, and Toyota Research Institute
Whether Generalist AI's GEN-1 99 percent reliability claim holds under third-party evaluation

FAQ: Physical AI Scaling Laws and What They Mean

Q: What exactly is a "scaling law for dexterity"?

A: NVIDIA's GR00T N1.7 research showed that increasing robot training data from 1,000 to 20,000 hours produces a predictable, measurable improvement in manipulation success rate. In language AI, scaling laws let researchers forecast model performance before training. The same predictability now applies to how well a robot can handle physical objects, which means labs can plan data collection roadmaps with confidence rather than guessing.

Q: How does GR00T N1.7 differ from earlier versions?

A: GR00T N1.7 uses an Action Cascade architecture: a vision-language model (Cosmos-Reason2-2B) generates action tokens, which a 32-layer diffusion transformer then converts into motor commands. Critically, it was trained on the EgoScale dataset of egocentric video, not expensive teleoperation data. The Apache 2.0 license means any company or researcher can use, modify, and deploy it commercially without restriction.

Q: Is BYD a serious humanoid robotics contender or just a press release?

A: The signals point to serious intent: the project started in 2022 (before the current hype cycle), 150 prototypes are inside BYD's own factories today, and the company has the battery expertise, supply chain, and global dealer network that most humanoid startups lack entirely. Whether BYD's timeline holds is an open question, but the underlying advantages are structural, not promotional.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

A Robot Got a Text and Walked to a Night Market in Taipei. Physical AI Just Left the Lab.

xBerry — Tue, 02 Jun 2026 07:21:47 +0000

Jensen Huang ended his GTC Taipei keynote with a robot navigating city streets autonomously to reach a night market. Amazon crossed 1 million robots. COMPUTEX 2026 declared "AI Goes Physical." Here is what this week means for Physical AI.

At the end of his GTC Taipei keynote, Jensen Huang showed a robot receiving a message about a party at the night market. Then the robot left - on its own, through city streets, to get there. No joystick. No remote operator. Just an agent with a destination and the physical capability to reach it.

It is a staged demo. Of course it is. But staged demos are how industries explain what they are building before the production version exists. And this particular demo, in Taipei this week, at the intersection of COMPUTEX and GTC, landed in a very specific way: the largest tech event in Asia just declared that Physical AI is no longer a laboratory concept. It is an agent navigating your city.

Metric	Value
Jensen Huang's humanoid market estimate	$40 trillion
Robots in Amazon warehouses, June 2026	1 million+
COMPUTEX 2026 exhibitors	1,500 from 33 countries
Jetson Thor vs Jetson Orin performance	7.5x more compute

Jensen Huang Put a Number on It. Then a Robot Walked Out the Door.

Jensen Huang called the humanoid robot market a $40 trillion opportunity at GTC Taipei. Wall Street responded with Physical AI stock moves before the keynote was over.

The number is large enough to invite skepticism, and it should. But the framing matters more than the precision: Huang is making the argument that humanoid robots will eventually address the same labor categories that humans currently fill across the global economy. That is not a 5-year claim. It is a 20-year structural argument.

The hardware that will get there is called NVIDIA Jetson Thor: 2,070 TFLOPs of FP4 compute, 7.5x more than Jetson Orin, designed specifically for on-device robot AI. The night market robot was not running on a server farm. It was running on something small enough to fit inside a humanoid chassis.

What the agentic framing means: When NVIDIA calls this "agentic AI going physical," they are describing a robot that receives a goal, plans a route, handles unexpected obstacles, and arrives. Not a robot that executes a pre-programmed path. The gap between those two things is the gap between industrial automation from 2010 and what Physical AI is building now.

COMPUTEX 2026 Declared Taiwan the Capital of Physical AI. Here Is Why That Matters.

The official theme of COMPUTEX 2026 is "AI Goes Physical". That is Taiwan's public statement about where it intends to sit in the next industrial order.

For decades, Taiwan dominated semiconductor manufacturing while largely leaving system integration and product design to others. COMPUTEX 2026, with 1,500 exhibitors from 33 countries across 6,000 booths, is the moment Taiwan signals it intends to move up the stack. Q1 2026 supply chain data already showed it: order books for humanoid actuators, gearboxes, and sensors from Taiwanese Tier 2 suppliers were growing faster than projections.

The geopolitical read is straightforward: the country that controls the physical AI supply chain - not just the chips, but the actuators, sensors, and integrated systems - will have structural leverage in the next decade the way semiconductor dominance provided leverage in the last one. Taiwan is not waiting to be assigned that role. It is claiming it.

Amazon Crossed 1 Million Robots. Nobody Made a Big Deal of It.

Somewhere in the past few weeks, Amazon crossed 1 million robots operating across its global warehouse network. There was no press release. No investor call highlight.

DeepFleet AI is improving routing efficiency across the entire network by 10%. The Sequoia system improved inventory identification and storage by 75% versus previous methods. One company is operating a robot workforce larger than the total warehouse labor force of most countries.

The reason this matters beyond the Amazon story: it proves the operational model at scale. The questions skeptics raise about humanoid robots - reliability, maintenance cycles, integration with existing workflows - Amazon has been answering these questions with non-humanoid robots for years. When Amazon moves seriously into humanoid deployment, they will not be running a pilot. They will be extending an existing operational competency.

NVIDIA Chose Unitree. That Is How Research Platforms Become Industry Standards.

NVIDIA selected Unitree H2 as the first commercial humanoid robotics system sold to research institutions - Stanford, ETH Zurich, and others. The package combines the 180cm Unitree H2 with NVIDIA Jetson Thor and the full Isaac software stack.

This is how research-to-industry pipelines get built. The models that Stanford researchers train on Unitree H2 this year will inform commercial deployments in 3 to 4 years. The companies whose hardware those researchers know intimately are the companies they will specify when the research becomes a product.

Unitree filed for IPO on Shanghai's STAR Board the same day, seeking 4.2 billion yuan ($620 million). The timing is deliberate: NVIDIA's endorsement lands on the same day as the public market application.

Starting Wednesday, CVPR 2026 in Denver runs the ManipArena Competition - the first benchmark evaluating AI models on 20 real manipulation tasks with actual robots, not simulators. The results will tell us which models actually work in the physical world. Watch that leaderboard.

What to Watch Next

ManipArena leaderboard at CVPR 2026 (June 3-7, Denver) - first honest comparison of which AI models actually work on real robots.
COMPUTEX Physical AI announcements through the week - big product reveals at 1,500-exhibitor events tend to come on days 2 and 3.
Unitree STAR Board IPO decision - a successful close would be a price signal for the entire sector.
Jetson Thor availability timeline - the shipping date determines when the research pipeline NVIDIA is building actually starts producing results.
Amazon Vulcan expansion - whether Vulcan's force-sensing capability extends beyond its current deployment will signal confidence in the dexterity problem being solved.

FAQ

Q: Is Jensen Huang's $40 trillion market claim realistic?

A: It depends entirely on the timeframe. Over a 20-30 year horizon, if humanoid robots reach the cost and reliability levels required to substitute for human labor across manufacturing, logistics, healthcare, and service industries, $40 trillion is a reasonable order-of-magnitude estimate. Over a 5-year horizon, it is not a useful number. The more relevant near-term figure is Bank of America's projection of 90,000 humanoids shipped in 2026 and 1.2 million by 2030.

Q: Why does NVIDIA choosing Unitree as a research platform matter for the broader market?

A: Research platforms become industry defaults. The hardware that PhD students and postdocs spend 4 years working with is the hardware they specify when they move into industry roles. NVIDIA selecting Unitree H2 for Stanford and ETH Zurich means the next generation of robotics engineers will have deep familiarity with Unitree hardware and the Isaac software stack. That institutional familiarity compounds into procurement decisions at scale over the following decade.

Q: What is the ManipArena Competition and why does it matter?

A: ManipArena, running at CVPR 2026 in Denver, is the first AI benchmark that evaluates models on 20 manipulation tasks using real physical robots rather than simulations. Simulation performance and real-world performance have historically diverged significantly. ManipArena results will be the most honest public ranking of which Physical AI models actually work. Watch the leaderboard: it will redirect research funding and commercial partnerships.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

79% of companies are already deploying Physical AI. Is yours one of them? Here's what you missed this week.

xBerry — Fri, 29 May 2026 08:29:34 +0000

Capgemini surveyed 1,700 organizations and 79% are already engaged with Physical AI. NVIDIA open-sourced GR00T N1.7 on Apache 2.0. Japan Airlines signed a 2-year deal at Haneda. Here is the week that moved Physical AI from pilot to platform.

A survey question was sent to 1,700 organizations across industries. The question was essentially: are you working with Physical AI yet? 79% said yes - not planning to, not evaluating it, but already engaged. 67% of CEOs in that same group called it a game-changer.

That is the Capgemini number from this week. And it lands differently than a funding headline or a robot demo. Funding can be speculative. Demos are controlled. A 1,700-company survey with 79% active engagement is a market temperature reading - and the temperature this week was unmistakably high.

Here is everything else that happened alongside it, and why the pieces fit together in a way that matters beyond each individual story.

Metric	Value
Companies already engaged with Physical AI	79% of 1,700 surveyed (Capgemini)
Training data for GR00T N1.7 EgoScale	20,854 hours of human POV video
Japan Airlines Haneda commitment	2-year operational deal with Unitree G1
GR00T N1.7 license	Apache 2.0 - fully commercial open source

79% Is Not a Forecast. It Is a Survey Result.

There is a meaningful difference between "X% of companies plan to adopt AI" and "X% of companies are already engaged." Planning is cheap. Engagement means teams, budgets, and at least one robot somewhere doing something real.

The Capgemini survey of 1,700 organizations found that 79% are already working with Physical AI and 67% of executives consider it a genuine strategic game-changer. BCG and Deloitte both published separate analyses this week reaching the same conclusion: the industry has crossed from a pilot phase into a strategy phase. These are not the same thing. Pilots have escape hatches. Strategy has budget lines.

Bessemer Venture Partners offered the most precise framing: this is the "GPT-2.5 moment" for robotics. Capabilities are real and demonstrably scaling. But the gap between current performance and the 99.9% production reliability required for full industrial deployment still exists. The analogy is useful because it tells you where we are on the curve: past the "does this work?" question, not yet at the "we can depend on this completely" answer.

What this means for your organization: If you are in the 21% not yet engaged, you are not safe - you are late. The companies currently running pilots are building institutional knowledge that compounds. The cost of catching up in 2027 will be higher than the cost of starting in 2026.

NVIDIA Just Open-Sourced the Brain of a Humanoid Robot. Here Is What That Changes.

On Friday, NVIDIA released Isaac GR00T N1.7 on Apache 2.0 - a fully commercial open-source vision-language-action model for humanoid robots. The license matters: Apache 2.0 means any company can use it in production, modify it, and ship products built on it without royalties or restrictions.

The technical story behind N1.7 is called EgoScale: NVIDIA pre-trained the model on 20,854 hours of video recorded from a human first-person perspective, covering 20+ task categories. From this, they derived the first observed scaling law for dexterity - increasing training data from 1,000 to 20,000 hours more than doubles manipulation accuracy. That is the same kind of predictable scaling that made large language models investable. When you can plot a curve and extrapolate it, you can plan a roadmap.

Humanoid, LG Electronics, and NEURA have already announced they are building on GR00T N1.7. Expect that list to grow fast. An open foundation model reduces the barrier for every robotics company that was previously spending resources on training from scratch. The gravitational effect is deliberate: NVIDIA is building the same ecosystem strategy for physical AI that it built with CUDA for GPU computing.

Why this matters: Open AI foundations accelerate the entire field. Companies that adopt GR00T N1.7 can focus engineering resources on application-layer differentiation rather than foundation model training. The cost curve for capable robots just dropped again - this time at the software layer.

Asia Is Moving Faster Than Western Boardrooms Realize.

Three separate Asia-Pacific moves this week tell a coherent story about who is treating Physical AI as infrastructure, not experiment.

Japan Airlines committed to a 2-year humanoid robot program at Haneda Airport using Unitree G1 robots (130cm, 35kg, 2-hour battery life). Tasks: baggage loading, cargo transport, cabin cleaning. Partner: GMO AI and Robotics. The driver is explicit - Japan's aging population is cutting labor availability while tourist traffic hits records. JAL is not deploying robots because it wants to. It is deploying because the demographic math leaves no alternative at scale.

Mitsubishi Electric and Chiba Institute of Technology signed a 3-year co-creation agreement to build Japan's own Physical AI stack from scratch: multi-legged walking robots, humanoids, and drones for infrastructure and emergency response. Mitsubishi brings precision motion control from its MELFA industrial robot line. Chiba brings large physics models for unpredictable environments. Japan is not licensing Physical AI from US companies - it is building sovereign capability.

In Singapore, IntBot and Certis Group announced a strategic partnership to deploy social robots in hotels, airports, hospitals, and shopping centers. Certis operates over 25,000 workers across Singapore, Australia, and Qatar. IntBot's layer is called General Social Intelligence - robots that recognize intent, hold conversations, and navigate crowded unpredictable spaces.

The pattern: Demographics, sovereignty, and dense urban environments are driving faster adoption in Asia than market analysis typically accounts for. The companies watching this from Western boardrooms should also be tracking the Taiwan supply chain data: Q1 2026 order books for humanoid actuators, gearboxes, and sensors are growing faster than projections across suppliers for Unitree, Figure AI, and 1X Technologies.

What to Watch Next

GR00T N1.7 adoption velocity - how many companies announce builds on Apache 2.0 in June will signal how fast the open ecosystem forms.
Robotics Summit Boston follow-through - the "State of Humanoids" panel had Boston Dynamics, Agility, and Schaeffler setting standards together. Watch for joint announcements in the weeks after.
JAL Haneda operational data - the first real performance data from a 2-year commercial airport deployment will be the most honest benchmark yet for humanoid reliability.
Capgemini 79% breakdown by industry - the aggregate is striking; the sector distribution will tell you which industries are leading and which are genuinely behind.
Physical Intelligence $1B close timeline - the round has been in negotiation; a close at $11B valuation resets the entire sector's comparable set.

FAQ

Q: What does "GPT-2.5 moment for robotics" mean in practice?

A: Bessemer's framing refers to the stage GPT-2.5 represented in language AI: capabilities were clearly real and scaling, but the technology was not yet reliable enough for most production use cases. For Physical AI in 2026, this means robots can handle structured tasks in controlled environments at meaningful scale, but the 99.9% reliability required for unsupervised industrial deployment is still a gap. The implication: invest and build now, because the reliability curve is predictable and the companies entering late will find the gap harder to close.

Q: Why does NVIDIA releasing GR00T N1.7 as open source matter for non-robotics companies?

A: Apache 2.0 means any company - including yours - can build a physical AI application on top of the same foundation model that Humanoid and LG Electronics are using, without licensing fees. The practical implication: the cost of building a capable task-specific robot application just dropped significantly. If your industry involves structured physical work, the barrier to prototyping a robotics solution in 2026 is lower than it has ever been.

Q: Why is Japan deploying humanoid robots faster than most other markets?

A: Japan has the most acute combination of demographic pressure and industrial precision culture of any major economy. The population is aging faster than any comparable country, labor availability in physical service roles is already constrained, and Japanese industrial culture has decades of comfort with robotics in manufacturing. Physical AI is not a disruption in Japan - it is a continuation of a 40-year automation trajectory, now applied to tasks that previous generations of robots could not handle.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.

A Startup founded in 2024 just signed a contract for thousands of robots at $25,000 each. Here Is the Moment Physical AI Made That Real.

xBerry — Tue, 26 May 2026 10:37:45 +0000

Humanoid prices fell from $85,000 to $25,000 in two years. Schaeffler signed a binding RaaS deal for thousands of robots starting December 2026. Hyundai's unions blocked 25,000 Atlas units. Physical AI's May 2026 reality check.

In 2023, a humanoid robot cost around $85,000. In 2025, that number dropped to $25,000, while profit margins actually improved. That is not a clearance sale. That is a technology cost curve doing what cost curves do when manufacturing volume compounds on top of model efficiency gains.

The question in 2023 was whether humanoid robots worked. In 2026, the question is different: who gets access first, at what price, and under what conditions. This week answered all three in ways that matter for every industry with structured physical work.

Metric	Value
Price drop	70% (from $85,000 to $25,000)
Hyundai Atlas plans	25,000 units from 2028 - blocked by union
VC in Physical AI 2026	$37 billion (all-time record)
Schaeffler RaaS start	December 2026

Humanoid Robots Crossed from Prototype to Commodity Pricing

The 70% price drop from $85,000 to $25,000 is the Unitree number, but it reflects a sector-wide dynamic. 1X Technologies began serial production of its NEO humanoid at its Hayward, California facility this month - the first US-based transition from R and D into repeatable factory output. Unitree is targeting 20,000 units shipped in 2026 after delivering 5,500 in 2025.

The Schaeffler deal is the most consequential signal of the week. A UK startup called Humanoid, founded in 2024 by Artem Sokolov, signed a binding contract with Schaeffler in mid-May. The model: Robot-as-a-Service (RaaS), with the first wheeled humanoid robots arriving at two German Schaeffler plants in December 2026. The target is thousands of units across Schaeffler's global facilities by 2032, with Schaeffler committing as a preferred actuator supplier delivering a 7-figure actuator volume by 2031.

A startup founded 18 months ago. A binding deployment contract. Thousands of robots. December 2026. That compressed timeline is the real story. The cost curve is not the only thing falling fast.

The Robot That Thinks Before It Moves, Not After It Fails

The standard model for robot learning has been: attempt, fail, correct, repeat. NVIDIA's Isaac GR00T N1.6, released this month, represents a different philosophy. It integrates NVIDIA Cosmos Reason - a slow-reasoning layer that makes the robot think through a task step by step before executing any physical movement. The system reasons explicitly before it acts, rather than learning from failure after.

Alongside GR00T N1.6, NVIDIA released Newton 1.0, a physics engine for dexterous manipulation, plus Isaac Sim 6.0 and Isaac Lab 3.0. The full training and validation stack is becoming an open platform.

Why this matters: If reasoning reduces manipulation errors before the motion happens rather than after, the quality bar for factory-grade robotics shifts significantly. Fewer failed attempts means fewer damaged products, fewer stoppages, fewer human interventions. For a manufacturer evaluating RaaS contracts, a reasoning robot is a fundamentally different risk calculation than a trial-and-error robot. Changing when reasoning happens - from post-action correction to pre-action planning - changes what robots can reliably commit to in production environments.

Where Real Friction Lives: 25,000 Robots and a Union Saying No

Hyundai announced plans to deploy 25,000 Boston Dynamics Atlas robots across its US manufacturing facilities from 2028, with initial operations at Metaplant America in Georgia handling parts sequencing. Hyundai is also building an actuator production facility targeting 350,000 units per year.

The Korean Metal Workers Union responded immediately: no Atlas robot enters any Hyundai plant without a labor agreement covering affected workers.

This is not an edge case. This is the playbook that will repeat in every country with organized labor and industrial robotics ambitions. The Hyundai situation maps the territory clearly: a company with capital, a confirmed technology, a deployment timeline, and a workforce with institutional leverage to negotiate terms.

The question is not whether unions can block deployment permanently. The question is what the negotiated terms look like: retraining commitments, transition timelines, revenue sharing, job guarantees in adjacent roles. Whoever gets this framework right first builds a deployment model others will follow.

The Capital That Makes This Permanent

Physical Intelligence is in negotiations for a $1 billion funding round. Mind Robotics closed $400 million. RoboStrategy listed on Nasdaq under ticker BOT as a public fund holding stakes in Figure AI, Apptronik, and Standard Bots. Total VC invested in Physical AI in 2026 has crossed $37 billion - a new all-time record with seven months still remaining in the year.

Barclays Research published "Robots roll out, economies rewire" on May 20. Key figures: humanoid robot market at $200 billion by 2035, China accounting for 85% of 2025 global deployments, robots potentially offsetting 60% of China's projected demographic workforce decline. The Barclays framing is the honest one. Not "robots will take jobs" but "economies will rewire." $37 billion in a single year is not speculative capital. It is directional commitment.

What to Watch Next

Humanoid+Schaeffler December deployment - the first binding RaaS contract at scale. If robots arrive on schedule in Germany, every Tier 1 supplier in Europe starts a new conversation.
Hyundai+union negotiations - the labor framework that emerges will be referenced by every industrial company deploying at scale.
GR00T N1.6 adoption rate - how many robotics companies build on the reasoning-first stack versus continuing with correction-based training.
Physical Intelligence $1B close - the valuation, reported at $11 billion, would reset comparables for the whole sector.
1X NEO throughput in Q3 - whether Hayward can sustain serial production is the US-based benchmark to watch.

FAQ

Q: What is RaaS and why does the Schaeffler deal matter?

A: Robot-as-a-Service means the customer pays per unit of work delivered, not for hardware ownership. Schaeffler does not buy robots outright - it pays for robot-hours in its factories. For companies evaluating humanoid adoption, RaaS removes the capital expenditure barrier and shifts risk to the robot provider. The Humanoid+Schaeffler deal matters because it is binding, names December 2026 as the start date, and the startup involved was founded in 2024. It is the clearest evidence that the RaaS model has moved from theoretical to contractual.

Q: What does GR00T N1.6 reasoning-first approach mean in practice?

A: Most current robots learn by doing tasks repeatedly, failing, and adjusting. GR00T N1.6 introduces slow reasoning: the robot works through the task plan step by step before any physical movement begins. In practice, this means fewer failed grasps, fewer product drops, fewer production line stops. For manufacturers in precision environments, this changes the reliability calculus significantly.

Q: Should workers in manufacturing be concerned about the Hyundai announcement?

A: The concern should be specific, not general. Hyundai's deployment begins with parts sequencing at Metaplant America in 2028. Tasks that are physically repetitive, dangerous, or high-precision are first. The Korean Metal Workers Union's response demonstrates that organized workforces have meaningful leverage to negotiate deployment terms. The question for workers is not whether robots arrive, but under what terms - and whether your workplace has a position before the contract is signed.

Physical AI Digest is a weekly briefing produced by Klaudia from xBerry - a tech company based in Poland building tools at the intersection of AI and operations.