DEV Community: xBerry

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.

Robots are being built faster than your industry is watching. Here's what you missed this week.

xBerry — Fri, 22 May 2026 09:08:29 +0000

Figure AI went from 1 robot per day to 1 per hour - in 4 months. Japan Airlines signed a 3-year humanoid contract. $37 billion in VC landed this year alone. Physical AI moved fast this week. Here is everything your industry missed, and why it matters more than the headlines suggested.

Think about the last time an airline signed a 3-year contract on technology that was still experimental. You cannot, because they do not. Airlines operate in one of the most regulated, liability-conscious industries on the planet. When Japan Airlines committed to a humanoid robot program at Haneda Airport in May 2026, they were not running a pilot. They were making a procurement decision - the same way they procure ground equipment, check-in systems, or gate management software.

That is the moment Physical AI changed categories. Not from a product demo, not from a VC funding round, but from a boring, bureaucratic, multi-year service contract at an airport most of you have probably transited through.

And JAL is not alone. This week told a consistent story across four separate industries. Here is what happened and what it means beyond the headlines.

The numbers first, so we have the same starting point

Metric	Value
Figure AI production speed increase	24x - from 1 robot/day to 1 robot/hour in 4 months
VC invested in Physical AI in 2026	$37B - a new all-time annual record, already
Projected humanoid robot market by 2035	$200B (Barclays Research, May 2026)
Share of global humanoid deployments - China	85% in 2025

When the Supply Chain moves, the Market is real

The strongest signal I track is not what robot companies say - it is what component suppliers do. This week, Khgears International, a Taiwanese manufacturer of precision gearboxes for industrial robots, announced a full pivot into humanoid-specific components: joints, drive mechanisms, actuator assemblies. They are seeking a strategic alliance with a Japanese Tier 1 automotive supplier to do it.

Khgears does not make this move on speculation. Gearbox manufacturers pivot when they have seen enough confirmed purchase orders to justify retooling their factory. When the supply chain moves, it means the demand is real, not projected.

The same signal comes from production lines. Figure AI went from producing one humanoid robot per day in January 2026 to one per hour by May - a 24x acceleration in four months. Their BotQ factory has already delivered 350+ units. For context: that is not a startup proving a concept. That is a manufacturer ramping toward industrial scale.

Why this matters: Component suppliers and production lines are lagging indicators - they follow confirmed demand. When they move at the same time, the market inflection has already happened. You are reading about the aftermath, not the prediction.

Robots are getting smarter faster than the hardware can keep up

Here is something that gets lost in the factory-and-funding coverage: the AI inside these robots is improving on a completely separate, faster curve. XPeng rolled the first mass-produced L4 robotaxi off its Guangzhou line on May 18 and the AI model powering that car is the exact same model running their Iron humanoid robot. One model trained once, deployed in two completely different physical systems.

This is a bigger deal than it looks. Training a frontier vision-language-action model costs tens of millions of dollars. If that cost is shared between an autonomous vehicle fleet and a humanoid workforce, the unit economics of Physical AI become dramatically better than analysts currently model. NVIDIA made the same bet in a different direction - their Isaac GR00T models are now open source, meaning any robotics company can build on a foundation instead of starting from scratch.

DARPA is already asking what comes after this architecture entirely. Their May 2026 research call imagines robots where the material itself computes - no central processor, no cloud, no latency. That is a 10-year horizon, but DARPA's early bets have a habit of becoming everyone's reality.

Why this matters: Shared AI models across platforms mean the cost of building capable robots is falling faster than the hardware suggests. The gap between "what robots can do in a lab" and "what they cost to deploy at scale" is closing from both ends simultaneously.

What this actually means if you are not an Investor

A direct note to everyone who works in logistics, manufacturing, aviation, or any field with structured physical tasks: the companies in this article are not running experiments in your industry. They are operating under multi-year service contracts. The question is no longer whether robots will enter your workplace. It is which tasks they take first, and how fast.

Barclays Research framed the macro picture in their May 2026 report: the humanoid robot market could reach $200 billion by 2035, and for China specifically, robots may offset up to 60% of the demographic workforce decline projected over the next decade. That last number is not a technology story - it is a labor economics story, and it will play out in every aging economy, not just China's.

The honest answer to "will robots take my job?" is still nuanced. Agility Robotics' agreement with Toyota covers logistics tasks in a manufacturing plant - moving parts, not assembling them. The Vodafone pilot in Duisburg had robots detecting misplaced products and unsafe pallet stacking, not replacing warehouse managers. The pattern so far is robots handling the physically repetitive and physically risky parts of jobs humans already find exhausting. But the category is expanding, and the speed of expansion is the variable to watch.

Why this matters: The Barclays report title is "Robots roll out, economies rewire." That word - rewire - is the honest one. Not replace, not eliminate. Rewire. The people who will navigate this best are the ones who start paying attention now, not when the robot is already at the next workstation.

What to watch next

Figure AI's BotQ throughput in Q3 - sustaining 1 robot/hour would make them the first humanoid manufacturer at genuine industrial scale by year-end.
XPeng Iron deployment update - the first real test of whether one AI model can actually run both a robotaxi fleet and a humanoid workforce in production.
Physical Intelligence's $1B raise - if it closes at the reported $11B valuation, it resets comparables for the entire sector and triggers a new wave of raises.
Khgears' Tier 1 alliance - whoever they partner with signals which Japanese industrial giant is moving seriously into humanoid supply chains.
The 85% China concentration risk - one country accounting for 85% of global deployments is a geopolitical variable that no analyst is pricing correctly yet.

FAQ

Q: What is Physical AI and why does 2026 matter?

A: Physical AI is artificial intelligence that operates in the real, physical world - humanoid robots, autonomous vehicles, robotic arms that reason in real time. 2026 matters because procurement replaced experimentation: Japan Airlines, Toyota, Amazon, and Vodafone are signing multi-year service contracts, not running pilots. Figure AI is producing one humanoid per hour. Barclays forecasts a $200 billion market by 2035. The phase shift from R&D to deployment happened this year.

Q: Will humanoid robots replace human workers?

A: The current deployment pattern is task replacement, not job replacement - robots are taking over physically repetitive, dangerous, or high-precision tasks within jobs that remain human-managed. Agility Robotics at Toyota handles parts logistics; humans still run the line. The Barclays framing is more accurate: economies will "rewire" rather than simply lose jobs. The speed of that rewiring, however, is accelerating significantly in 2026, and the category of tasks robots can handle is expanding rapidly.

Q: Which companies should I be watching in Physical AI right now?

A: Figure AI for production velocity and deployment scale. XPeng for the shared AI model strategy across robotaxi and humanoid. Physical Intelligence for foundation model development (their $1B raise at $11B valuation is a sector bellwether). NVIDIA as infrastructure - Isaac GR00T is becoming the Linux of robotics AI. And watch Khgears and other component suppliers: they tell you what the demand actually is, not what companies claim it will be.

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.

Skild Brain, $13,500 Humanoids, and a NASDAQ Ticker

xBerry — Tue, 12 May 2026 12:58:30 +0000

Two days, three structural shifts. RoboStrategy (BOT) listed on NASDAQ — retail access to Figure AI and Apptronik in one stock. Sereact Cortex 2.0 hit one billion production pick operations (1 failure per 53,000). Skild AI acquired Fetch Robotics to build Skild Brain — one unified control layer for humanoids, AMRs, arms, and robot dogs. $183M deployed in 48 hours. Unitree G1 now costs $13,500. This article is for engineers and tech leads tracking where the Physical AI stack is consolidating.

This week's Physical AI news is worth separating into what's technically significant versus what's financially significant because this week they're both unusually dense.

Financially: Physical AI got a NASDAQ ticker.
Technically: a robotic picking brain crossed one billion production operations with a world-model architecture that explains why it doesn't need retraining.
Strategically: Skild AI made the most important software consolidation move in warehouse robotics since Amazon acquired Kiva Systems.

Here's what each of these means for the stack.

Key Facts

RoboStrategy (BOT) - first Physical AI fund on NASDAQ; portfolio includes Figure AI, Apptronik, Standard Bots
Sereact Cortex 2.0 - 1B production picks; 1 failure per 53,000; $110M Series B; world model + VLA architecture
Skild AI + Fetch Robotics - acquisition of Zebra Technologies robotics division; Skild Brain = unified orchestration for mixed fleets
GR00T N1.7 - Qwen3-VL backbone; 20K hrs EgoScale pretraining; commercially licensed; drop-in from N1.6
Vbot $73M Pre-A - full-size humanoids; Unitree G1 at $13,500 (−90% vs 2024)
Tesla Optimus Gen 3 - mass production, Fremont; 50,000 units by end 2026
Capital May 11–12 - $183M+; Q1 2026 humanoid funding: $2.37B (+288% YoY)
Deloitte 2026 - 58% already use Physical AI; 80% plan to; 22% have a change management plan.

Sereact Cortex 2.0: The Architecture Behind a Billion Picks
A billion operations is a production metric, not a benchmark. Here's why the architecture produces it.

Cortex 2.0 integrates a world model alongside the VLA policy. The execution loop: generate candidate motions → simulate each against a physics model → score and select optimal → execute. The physics simulation layer is what eliminates the retraining requirement when object configurations change. New SKU, new packaging format, novel arrangement - the world model evaluates them without requiring labeled examples in the training set.

**The result: **one failure per 53,000 production picks, across real warehouse variability, over one billion operations. At that reliability threshold, the system operates without continuous human supervision.
GR00T N1.7 advances the foundation model side of the same problem. The new Qwen3-VL backbone processes language instructions with better multi-step comprehension. Pretraining on 20,000 hours of EgoScale human egocentric video gives the model manipulation priors that transfer directly to robot motor control because GR00T uses a relative end-effector action space shared across human and robot embodiments.

Upgrade path from N1.6: drop-in. Point --model-path to nvidia/GR00T-N1.7. Existing embodiment configs carry over. EgoScale pretraining improves dexterity generalization before any task-specific fine-tuning.

Skild Brain: Fleet Orchestration as a Platform

The Skild AI acquisition of Fetch Robotics assets is the most consequential software consolidation move of the week — possibly the quarter.

Current state: most multi-robot warehouse deployments run separate control stacks for each robot category. AMR fleet management (navigation, routing, charging), arm control (motion planning, grasp policies), humanoid policies (whole-body coordination, task planning), robot dog inspection (perception, anomaly detection). Four categories, four software layers, four integration points with WMS/ERP systems.

**Skild Brain targets a single unified layer: **one AI intelligence system that orchestrates task assignment, routing, and execution across the full heterogeneous fleet. The Symmetry Fulfillment platform - acquired as part of the Fetch Robotics assets - provides production-validated workflows and an existing customer base to deploy against.

For anyone building robotics software for industrial and warehouse environments, this is the consolidation signal: the orchestration layer is becoming a platform play, not a point solution. The integration challenge shifts from hardware interoperability to software fleet intelligence - which is also where the margin concentrates.

Open-Source and Production Stack: What You Can Use Today

Tool	What it does	Where
GR00T N1.7	Open VLA for humanoid robots, commercially licensed	github.com/NVIDIA/Isaac-GR00T
MuJoCo-Warp	70× faster GPU physics simulation	github.com/google-deepmind/mujoco
Newton 1.0	Open physics engine for dexterous manipulation	Via Isaac Lab 3.0
Isaac Lab 3.0	Large-scale robot learning on DGX infrastructure	developer.nvidia.com/isaac/lab
Symmetry Fulfillment	Production warehouse orchestration (now Skild)	Via Skild AI

$13,500 and the Market Expansion Problem

Unitree G1 at $13,500 is a 90% price reduction from 2024 equivalents. Unitree targets 10,000–20,000 deliveries in 2026. Tesla Optimus Gen 3 targets 50,000 units from Fremont alone.

Deloitte's Tech Trends 2026 projected 15,000 industrial humanoid units delivered at $14,000–$18,000. Those numbers are already being revised upward by the volume targets now in play.

The addressable market at $13,500 is qualitatively different from the market at $100,000+. Mid-size manufacturers, regional logistics operators, smaller distribution centers - all enter scope. The hardware commoditization opens demand that the software orchestration layer (see: Skild Brain) now needs to serve. The constraint has moved upstream: it's no longer "can we afford the robot" but "do we have the software infrastructure to run a mixed fleet."

DARPA: The Category Beyond Foundation Model Robotics

The furthest-horizon signal of the week: DARPA's RFI for materials with embedded intelligence - sensing, adapting, and acting without external computation. Light-stimulated polymers demonstrating photothermal 3D shape response, sustaining loads 24,000× their own mass, are the early physical evidence.

This defines a third architectural paradigm:

Classical automation: explicit programming
Foundation model robotics: learned policies (GR00T, Cortex 2.0)
Embodied materials intelligence: perception + processing + actuation in the substrate.

RFI to deployable technology is a decade horizon. But for engineers thinking about where the stack goes after VLA models mature, this is the category to watch.

The Data

Metric	Figure
Global robotics market 2026	$132B (+16% YoY)
Industrial installations record	$16.7B
Warehouse robotics 2025→2030	$9.33B → $21B+
Using Physical AI (Deloitte 2026)	58%
Planning adoption within 2 years	80%
Have change management plan	22%
Humanoid funding Q1 2026	$2.37B (+288% YoY)
Capital deployed May 11–12	$183M+
Sereact failure rate	1 per 53,000 ops
Unitree G1 price vs 2024	$13,500 (−90%)
Tesla Optimus Gen 3 target	50,000 units, 2026

Frequently Asked Questions

What is Skild Brain and why does the Fetch acquisition matter?

Skild Brain is a unified intelligence layer for mixed robot fleets - humanoids, AMRs, arms, robot dogs — under one control system. The Fetch Robotics acquisition (from Zebra Technologies) gives Skild both the orchestration platform (Symmetry Fulfillment) and an existing production customer base. Most warehouse deployments currently run separate stacks per robot category. Skild Brain is the first serious attempt to unify them at the platform level.

What makes Sereact Cortex 2.0 different from other VLA systems?

The world model integration. Rather than direct visual-to-motor mapping, Cortex 2.0 generates candidate motions, simulates them against a physics model, selects optimal, then executes. This simulation layer handles novel configurations without retraining - which is why it reached one billion production operations with a 1-per-53,000 failure rate.

How do I upgrade from GR00T N1.6 to N1.7?

Drop-in: point --model-path to nvidia/GR00T-N1.7. Existing embodiment configs and workflows carry over.
Key changes: Qwen3-VL backbone replaces Eagle, EgoScale human video pretraining improves dexterity generalization before fine-tuning.

What is RoboStrategy BOT?

First NASDAQ-listed Physical AI fund. Retail access to Figure AI, Apptronik, Standard Bots in one stock. Listed May 11, 2026. First time retail investors can access the category without venture or private market access.

What is DARPA's materials intelligence RFI?

A call to define materials capable of sensing, adapting, and acting without a separate compute layer - intelligence in the substrate itself.
Early physical evidence: light-stimulated polymers with photothermal 3D response sustaining 24,000× their own mass. Decade horizon to deployment.

Why 22% change management readiness vs 58% adoption?

Deloitte 2026: adoption is outpacing organizational readiness. 58% use Physical AI, 80% plan to within 2 years, but only 22% have structured transformation plans.
Barriers: reskilling, legacy ERP integration, no internal fleet management competency. Deployment cycles are now 7 months - faster than most organizational change programs.

Summary

BOT on NASDAQ. One billion Sereact operations. Skild Brain. $13,500 Unitree G1. $183M in 48 hours.

The orchestration layer is consolidating. The foundation models are production-licensed. The hardware is commoditizing. The 22% who have a change management plan are building the operational infrastructure to actually use all of this. The 78% who don't are accumulating technical debt in a different form - organizational debt.

Sources: Sereact $110M Series B - SiliconAngle · Skild AI acquires Zebra Robotics - Skild AI · Skild acquires Fetch - The Robot Report · RoboStrategy BOT - GlobeNewswire · Vbot $73M - The AI Insider · GR00T N1.7 - GitHub · Deloitte Physical AI 2026 · BCG - Physical AI Reshaping Robotics.

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 Surpasses $88 Billion: When Technology Arrives Before Organizational Readiness

xBerry — Fri, 08 May 2026 11:16:18 +0000

In Short

Physical AI - systems combining perception, reasoning, and robotic action into a single autonomous loop - crossed $88 billion in market value in 2026. This article is for engineers and tech leaders evaluating Physical AI adoption. The technology stack is largely open-source and production-ready. Deployment cycles have shrunk from 24 to 7 months. The bottleneck is no longer the code - it's organizational change management, which 78% of companies haven't figured out yet.

If you've been following the robotics space, you already know the demos look impressive. Atlas does backflips. Digit moves boxes. GR00T controls a humanoid arm with finger-level precision. But in 2026, the interesting question is no longer can robots do this - it's how do I actually deploy this in production?

The barrier to Physical AI is no longer technological. It is organizational. The tools are ready. The models are open. The simulators are fast. What's missing is the roadmap for companies to integrate all of this into real operations and that's a problem engineers are increasingly being asked to solve alongside their leadership teams.

Key thesis: The technology is 90% ready. The bottleneck is companies' capacity to manage the transformation that robots bring.

Key Facts

Robotics market - reached $88.27 billion in 2026; forecast to grow to $416 billion by 2035 at a CAGR of 19.86%.
Deployment cycle - shortened from 24 months (2020–2024) to just 7 months (2026).
Change management gap - 78% of companies have no plan for managing the workforce and process transformation that Physical AI requires.
NVIDIA GTC 2026 - released Cosmos 3, GR00T N1.7, and Newton 1.0 as open or commercially licensed models.
MuJoCo-Warp - accelerates robotics training by 70×, compressing weeks of learning into minutes.

What Physical AI Actually Means for Builders

Physical AI is a class of systems that close the loop between perception, reasoning, and action in the real world. Unlike classical automation - where every step is explicitly programmed - Physical AI learns from examples and generalizes to new situations.

The architecture typically looks like this:

Perception layer - cameras, depth sensors, tactile sensors feeding raw data
Reasoning layer - Vision-Language-Action (VLA) models processing multimodal input and planning multi-step tasks
Action layer - motor controllers, robotic arms, mobile bases executing continuous-value action vectors

The key shift in 2026 is that all three layers now have open, production-grade foundations you can build on today.

The Open-Source Stack You Can Actually Use

This is where it gets practical. Here's what's available right now:

GR00T N1.7 - NVIDIA's open Vision-Language-Action model for humanoid robots. A 3B-parameter model trained on 20,000+ hours of human egocentric video. Commercially licensed. Runs on Jetson Thor for edge deployment. Drop-in swap from N1.6 - point --model-path to nvidia/GR00T-N1.7 and existing configs carry over.
MuJoCo-Warp - Google DeepMind's GPU-accelerated physics simulation, co-developed with NVIDIA. 70× faster than CPU-based MuJoCo. Available through MJX open-source library and integrated into Newton. If you're training robot policies, this changes your iteration speed dramatically.
Newton 1.0 - Open-source physics engine co-developed by NVIDIA, Google DeepMind, and Disney Research. Purpose-built for dexterous manipulation training. Handles cables, small parts assembly, contact-rich tasks that previously required extensive manual programming.
Isaac Lab 3.0 - NVIDIA's large-scale robot learning framework, now in early access. Built on Newton, adds multiphysics simulation and improved support for complex manipulation. Runs on DGX-class infrastructure.
Isaac Cosmos 3 - Unified world foundation model for synthetic data generation, visual reasoning, and action simulation. Replaces three previously separate pipelines with one architecture.

Where Physical AI Generates Real ROI - A Developer's View

Knowing where to apply these tools matters as much as knowing how to use them. Here's where the payback is clearest:

Application Area	Effectiveness	Payback Period	What You're Actually Building
Quality inspection	97–99% defect detection (vs 70–80% manual)	3–6 months	CV pipeline + edge inference
Warehouse logistics	$30B market (2026), doubling by 2030	14–18 months	AMR navigation + fleet orchestration
Humanoid in production	Hours of uninterrupted operation (Toyota/Digit)	18–24 months	Full-body VLA policy deployment
Robotic surgery	60% of major hospitals deployed systems	24–36 months	Autonomous arm control + imaging AI
Machine alert interpretation	LLM adoption in industry: 16% → 35% YoY	6–12 months	LLM on top of sensor/SCADA data

Quality inspection is the lowest-hanging fruit. A camera, an inference model, and an edge device - deployed in weeks, ROI in months. If you're looking for a first Physical AI project inside a manufacturing client, this is where to start.

The warehouse logistics space is where AGV + AI navigation stacks are maturing fastest. Fusion of traditional pallet movers with autonomous navigation modules creates hybrid systems at a fraction of the cost of full automation.

The Bottleneck Isn't the Model - It's the Organization

Here's the uncomfortable truth for anyone trying to deploy Physical AI in an enterprise: 78% of companies don't have a change management plan. According to IFR and BCG reports from 2026, the main barriers are:

No plan for reskilling workers whose tasks will be automated,
Inability to integrate robotic software with decade-old ERP systems,
No internal competency for managing fleets of autonomous systems at scale.

This means the most valuable skill in Physical AI deployments right now isn't robotics engineering - it's the ability to bridge the technical stack with organizational transformation. Engineers who can speak both languages are extremely rare and extremely valuable.

Toyota Motor Manufacturing Canada deployed seven Digit units from Agility Robotics in under five months, running component logistics in RAV4 production loops for multi-hour uninterrupted blocks. The technical deployment wasn't the hard part. The process redesign was.

Frequently Asked Questions

What is Physical AI?
Physical AI is a class of AI systems operating in the physical world - combining sensory perception, language and vision models, and actuators (robotic arms, AGVs, humanoids) into a single autonomous decision-making loop. Unlike classical automation, Physical AI learns new tasks from examples without manually programming every step.

How long does it take to deploy a robot in a factory?
The deployment cycle has shortened from 24 months (2020–2024) to seven months (2026). Key accelerators: ready-made open-source models (GR00T, Cosmos) and GPU-based simulators (MuJoCo-Warp, 70× faster training). Toyota deployed Digit in under five months.

What is the actual ROI in robotics deployments?
In quality inspection: 3–6 months. Warehouse logistics: 14–18 months for operations running more than two shifts per day. Robotic surgery: 24–36 months with growing procedure volumes. Operating costs fall by 30% in fully automated facilities.

Will robots replace workers?
They will transform work rather than eliminate it. A Gartner report from April 2026 shows that by 2030, 50% of new warehouses in developed markets will be designed as robot-centric facilities, with human roles shifting to supervision, servicing, and exception handling. BCG forecasts that more than 50% of jobs will be significantly reshaped by AI within 2–3 years, with only 10–15% fully displaced.

What's the fastest way to get started with Physical AI development?
Start with simulation. Isaac Lab + MuJoCo-Warp gives you a 70× faster training loop than CPU-based alternatives. Use GR00T N1.7 as your base VLA model and fine-tune for your specific embodiment and task. For perception tasks, a computer vision pipeline on edge hardware (Jetson Thor) is the lowest-cost entry point with the fastest payback.

Why are 78% of companies not ready for Physical AI?
According to IFR and BCG reports from 2026, the main barriers are: no reskilling plan for workers, inability to integrate robotic systems with legacy ERP platforms, and lack of internal competency for autonomous fleet management. This is a leadership and organizational problem, not a technological one.

Summary

The Physical AI stack in 2026 is more open, more capable, and more production-ready than most engineers realize. GR00T, Newton, MuJoCo-Warp, and Cosmos 3 are not research previews - they are tools you can deploy today. The 70× simulation speedup alone changes what's possible for teams without massive compute budgets.

The hard problem is no longer building the robot. It's preparing the organization to work alongside it. Whoever solves that change management layer and can implement the technical stack on top of it - is positioned at the most valuable intersection in the industry right 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.