Kawasaki, FANUC, and Yaskawa pooled $12.3M from NEDO to build a shared tactile robot foundation model, while DeepMind's Gemini Robotics 2 gives humanoids whole-body control that adapts to new robot bodies in hours, pushing platform-level robot AI forward on two fronts at once. Meanwhile the FCC banned foreign-made networked robots over 4.4 lb from US sale on cybersecurity grounds, directly blocking Unitree's H1 Pro launch just as Unitree priced its Shanghai IPO at a $9B valuation with DeepSeek buying in.
What happened 19
The Robot Report · 2026-08-08
Frames where hardware and robotics deployment is heading, from Brooks' industrial-robot shipping history to Hurst's loco-manipulation work at Agility.
“Brooks and his students built the first mobile robots that moved among people at human speeds, then the first humanoid robots that used AI computer vision and simple speech understanding to interact with people.”
“His later companies have delivered tens of millions of Roombas into people's homes, thousands of ground robots to the U.S. military in Iraq, and thousands of upper-torso humanoid robots to factories worldwide.”
“Currently, Hurst works on physical AI to facilitate complex loco-manipulation behaviors, exemplified by Digit, the first commercially deployed humanoid robot in logistics.”
“Goldberg is the director of the AUTOLab and CITRIS 'People and Robots' Initiative at UC Berkeley.”
The Humanoid Hub · 2026-08-08
First time direct competitors in industrial robotics pool training data — a sign no single manufacturer has enough data alone for a competitive foundation model.
“The goal is a VTLA foundation model built from 5,000 hours of visual, tactile, and motor data from real manufacturing environments.”
“Kawasaki, FANUC, and Yaskawa are direct competitors in the industrial robotics market.”
“Individually, no manufacturer has enough data volume for a capable foundation model — together, their installed bases exceed critical thresholds.”
“Japan has set a goal of controlling around 30 percent of the global physical-AI market by 2040.”
The Robot Report · 2026-08-07
Dexterous, tactile-sensing hands are the missing piece for robots to do skilled manual work, not just pick-and-place.
“We have vision-only robots and a variety of different robots that do really good work today, which is moving objects and so on, but the holy grail is to be able to augment, complement, and expand the potential to use robots to do high-value work.”
“For high-value work, we need our robotic hands to be able to touch and feel the real world.”
“That allows us to make the fingers very, very thin.”
“These are not MEMS-based sensors — each one has its own analog-to-digital conversion, and the scale, the pitch, is 1.”
Dwarkesh Patel · 2026-08-07
Argues that current session-bound LLMs are structurally capped until they can accumulate experience the way humans do, which bears directly on how fast embodied AI can actually get good at real tasks.
“I don't think you can have AIs that perform whole jobs as competently as humans if they are forced to just write markdown files from session to session.” 0:00
“I don't think there's any sequence of text they could write to each other that would allow the subsequent student to just nail the saxophone on the first try.” 0:35
“I don't think this assumption necessarily makes sense in the future, and this is one of the many reasons I'm worried about locking in some kind of safety regulatory regime right now.” 1:15
“What if the model is improving every single day based on the millions of sessions of work it does that day?” 1:15
Robohub · 2026-08-07
Automating simulation-content generation attacks the data bottleneck that's held back general-purpose robot learning.
“One natural idea is to use simulation as a training ground. While there has been significant progress over the last few years in the physics engines that power robotics simulators, one of the remaining challenges has been creating sufficiently rich and diverse simulation content to capture the complexity of the real world.”
“First, a "designer" agent generates the elements of a scene, then a "critic" advises whether it looks realistic, and finally, an "orchestrator" manages their back-and-forth, deciding when the design is done.”
Latent Space · 2026-08-07
A major chipmaker paying to own model-specific inference silicon signals custom ASICs, not general GPUs, as the next compute battleground.
“In The Custom ASIC Thesis we said Taalas was worth paying attention to, and in the Inference Inflection we said everything would go vertical.”
“We built Taalas to rethink AI inference from the ground up: hardware designed around the model, rather than the other way around.”
“The result is the world's fastest and most cost-effective inference silicon.”
2026-08-07
If pretraining data is the real ceiling, scaling compute alone won't keep frontier models improving at the same pace.
“Compute is growing, but the data is not growing, because we have but one internet.” 1:56
“The network has to compress the world behind the text — the facts, the styles, the reasoning.” 1:10
“The text is a projection of the world, and the model learns the process that produced it.” 1:33
“What Ilya saw was not a demo, it was a limit — a ceiling built into the way every modern AI is made.” 0:02
The Robot Report · 2026-08-06
A defence prime committing near-billion-dollar, milestone-gated production volume to two robotics startups is the clearest signal yet that physical AI has a real industrial buyer.
“HII intends to award up to $900 million in total shipbuilding work to Path Robotics and GrayMatter.”
“The funding is contingent on the two companies meeting clearly defined technology and manufacturing readiness, as well as performance milestones, said HII.”
“Together, we are defining a new approach to robotics in shipbuilding: automation that can adapt to extreme levels of complexity, mix, and size.”
Robohub · 2026-08-06
It cuts US startups off from cheap Chinese platforms and hands allied and European manufacturers a protected window to win share.
“The block applies to networked humanoids, quadrupeds, other qualifying mobile robots weighing more than 4.4 lb.”
“In the near term, the measure could slow U.S. physical AI innovation by cutting startups and researchers off from future low-cost Chinese platforms before comparable Western alternatives exist.”
“Yet excluding foreign products is not the same as building a competitive industrial base.”
“On July 30, Reuters reported that China’s Commerce Ministry threatened to “resolutely retaliate” if the United States maintains the restrictions and urged the United States to immediately withdraw the ban.”
a16z · 2026-08-05
Ocean robotics has crossed from demand creation to production scaling, pulled by AI-driven demand for subsea cables, minerals, and energy.
“We wanted to massively scale up underwater ecosystem restoration. But in doing that and going after that problem, we realized that there is a lot more than underwater ecosystem restoration that needs to be scaled up.” 0:51
“It's increasing demand for critical minerals. It's increasing demand for energy.” 1:14
“The demand hasn't been constrained to one domain like defense or commercial crowd.” 1:38
“The company is called Ulysus, and it says it's created drones that not only can detect these mines, but more importantly can disarm them.” 2:31
The Humanoid Hub · 2026-08-05
Teleoperation-first deployment generates revenue and training data simultaneously, sidestepping the wait for full humanoid autonomy.
“Humanoid robots are controlled by human remote operators from day one — this allows them to be deployed productively immediately, without having to wait for mature autonomy software.”
“Every interaction simultaneously generates training data that continuously improves the in-house robotics foundation model and gradually increases the share of AI-controlled operations.”
“Since December 2025, the robots have packed, sorted and prepared more than 900,000 products for delivery in real customer environments.”
The Robot Report · 2026-08-04
Shifts the humanoid bottleneck from a research problem to a cost-per-unit problem, which changes what a hardware investor should underwrite.
“While the tech industry thinks AI is the main barrier for robots joining the manufacturing workforce, the real issue has less to do with intelligence and more to do with robotic unit economics and financial viability.”
“The conversation has moved from “Can they do the work?” to “Can we justify the cost?”
“Some developers lean hard on teleoperation, where a person controls the robot’s movements directly. That requires almost all of the operator’s attention, which caps how far it can scale because every robot needs a human tied to it.”
“A better model is to treat people as high-level supervisors who step in only when a robot faces something unusual or needs direction on a task. This way, one person can supervise 50 or more robots at once, and every intervention serves as training data to improve the AI over time.”
“A factory automation system built for one workflow is optimized for it from end to end. In high-volume production, that specialization beats general-purpose humanoids.”
The Humanoid Hub · 2026-08-04
The largest EV maker is entering humanoids using its own battery, motor and control supply chain, which reprices the component advantage pure-play robotics startups have to buy.
“The 1.61-meter-tall robot speaks six Chinese dialects and six foreign languages and is to advise customers in BYD dealerships in the future — a direct strategic counter to Tesla Optimus.”
“BYD's Executive Vice President Stella Li had already outlined the goal: two to three robots per showroom that present vehicles, demonstrate features and greet customers.”
“The company has years of expertise in developing battery systems, electric motors and control software — core components that are also needed in humanoids. Pure robotics companies have to build up this supply chain; BYD can source internally.”
“While the FCC ban excludes Unitree and other Chinese manufacturers from the US market, BYD as a car company has no US sales yet — and therefore loses less through the regulation than the pure robotics companies.”
Don't Worry About the Vase · 2026-08-03
Machine-verified solutions to long-standing open problems at trivial compute cost marks a step change in what frontier models can do in genuinely novel research.
“The results were achieved by an internal version of Astra, our next major model. The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates.”
“These arguments were then prepared into manuscripts by humans with the same model. Afterward, the model formalized each argument in a Lean certificate.”
“Connes’s rigidity conjecture. Disproof of a longstanding conjecture that certain groups are uniquely determined by their von Neumann algebras.”
“And yes we did try other major problems without success. Sadly no Millennium Prize problems (yet).”
“But also, we didn’t spend a lot on each problem. It’s possible to push test-time compute much further.”
Dwarkesh Patel · 2026-08-03
If compute prices rise with capability, every robotics and embodied-AI company modelling falling inference costs into its margins is planning against the wrong curve.
“For the last three consecutive years, Anthropic's revenue has 10x'd year over year, and it's likely to do so again this year. They ended last year with nine billion in revenue.” 0:00
“For a lab to keep 10x'ing revenue year over year while compute only 3x's, one of the following three things needs to happen, or some combination of the three: One, lab margins have to increase. Two, the price of compute has to increase. Or three, the percentage of compute that labs spend on inference rather than training has to increase.” 0:39
“Anthropic's inference margins reportedly went from forty percent in the middle of last year to upwards of eighty percent now for Fable.” 1:01
“The spot prices for compute are more than forty percent higher than they were in the February trough that we had earlier this year.” 1:01
“In 2024, according to Epoch, OpenAI was spending just a quarter of its compute on inference, and that number is likely closer to fifty percent, if not higher, now.” 1:20
“If you're spending most of your compute on inference, then you're basically declaring that AI progress has stalled and you're just now in the business of being a cloud provider.” 1:37
The Robot Report · 2026-08-03
A London-based DeepMind applied-robotics spinout betting on on-site human teaching instead of programmers, already deployed and pre-seed funded by European VCs.
“A useful robot should be able to learn from the person doing the work,.”
“They should be able to show it a task, put it right when it makes a mistake, and move on to the next problem.”
“He later co-founded Reimagine Robotics in April 2025 alongside former colleagues Oleg Sushkov, Akhil Raju, and Misha Denil, with headquarters in London and Sydney.”
“First phase was funded by pre-seed financing from Fly Ventures, firstminute capital, and a number of angel investors.”
“At a made-to-order plastics business, Reimagine Robotics trained its robots to tend 3D printers overnight by removing print beds, operating latches, and pressing controls.”
The Humanoid Hub · 2026-08-03
US market closure pushes the highest-volume Chinese humanoid maker onto Europe, exactly where EU AI Act conformity duties just took effect.
“On July 28, 2026 — six days after the Europe launch — the US regulator FCC put humanoid robots from foreign manufacturers on its national ban list.”
“The FCC ban applies exclusively to new models and approvals — devices already approved before remain legal.”
“No European company offers a comparable humanoid in comparable volume, and the shortage of skilled workers in European manufacturing industry creates real demand for automation solutions.”
“The EU AI Act prescribes mandatory conformity requirements for high-risk AI systems — and autonomous humanoids in industrial environments could fall under this — which took effect on August 2, 2026, eleven days after the H1 Pro's Europe launch on July 22.”
The Robot Report · 2026-08-02
Whole-body VLA control plus few-hour cross-embodiment transfer moves the bottleneck for humanoid startups from motion stacks to data and hardware.
“The model can now control entire humanoid robots, translating intent into intelligent whole-body control.”
“Gemini Robotics 2 can also run locally on-device while adapting to entirely new robotic bodies in just a few hours, claimed DeepMind.”
“Apollo processes the instruction, walks to the table, and picks up the watering can, takes a few steps to the shelves, and places it precisely in its destination.”
“Enabling robots to communicate with people, understand the physical world, and plan multi-step tasks lasting several minutes.”
The Humanoid Hub · 2026-08-01
Ladder climbing is the hand-foot coordination test that separates warehouse demos from real facility inspection and maintenance work.
“Figure hat weder technische Details zum Helix-System noch zur Trainingsmethodik offengelegt.”
“According to Figure, the company's own Helix AI system controls the process — the same system also used in the BMW deployments for sequencing tasks.”
“Ladder climbing is considered one of the more demanding mobility tests for humanoid robots: the task requires simultaneous coordination of hands and feet, continuous balance adjustment and precise spatial perception.”
“Adcock admitted that the robot's current climbing speed is "methodical" — that is: slow.”
What people said 6
Guillaume Verdon · 2026-08-06
If thermodynamic computing scales as claimed, it could sidestep the power wall limiting classical compute growth — a direct physics-of-computation bet.
“We've scaled up our chips quite a bit. We've been scaling a thousand X year over year.” 0:01
“You can do all sorts of AI algorithms with about 10,000 times fewer transistors than if you were to do them digitally.” 2:51
“Each transistor is operating in a regime that would usually be considered janky.” 3:33
“We've been working on reinventing computing from the ground up for a couple of years now.” 1:17
Gavin Baker · 2026-08-06
Compute demand and rental pricing kept climbing through the selloff, and Nvidia is now financing its own customers' capex.
“Every set of quantitative indicators he tracks is accelerating — GPU rental prices up 50%-60%, hyperscale cloud providers' operating cash flow growth climbing from 28% to 32%. There is a significant disconnect between market sentiment and the actual data on AI compute demand.”
“The market is conflating 'a shift in the AI competitive landscape' with 'a collapse in AI investment.'”
“But Baker believes the market's sell-off of Nvidia has already become excessive, with its forward price-to-earnings ratio falling to near its lowest level this year. While some companies are developing their own AI chips, they remain highly dependent on Nvidia's technology, and there is no sign yet that Nvidia will lose its industry leadership.”
“This model allows AI cloud providers to deploy Nvidia infrastructure at scale and offer compute services externally without having to bear the full upfront capital expenditure. In exchange, in addition to revenue from chip sales, Nvidia will also take a percentage cut of the cloud providers' revenue.”
Karl Friston · 2026-08-04
Friston frames the free-energy account of mind as continuous with older physics of sense-making, the theoretical root of active inference.
“What makes us tick and what can go wrong when we when we don't make the right kind of sense making or decision making.” 0:55
a16z · Fei-Fei Li · 2026-07-28
World Labs is betting that generated worlds replace real-world data collection and evaluation for robotics.
“Spatial intelligence is about creating AI that has the ability to generate understand reason with and interact with spaces whether it's physical or virtual. And of course a means to an end towards spatial intelligence is building large world models.” 1:26
“I mean, look at the creative field, whether it's VFX or gaming and or design. Many use cases, you can create and act within virtual spaces.” 2:31
“World Lab has is a two-year-old startup. I think we should just recognize it's a frontier model lab.” 0:51
Anil Seth · 2026-07-28
The predictive-processing account of perception is the same generative-model framing that world models and active inference build on.
“We don't passively perceive the world, we actively generate it.” 1:45
“And it's very easy to take for granted that this world just pours itself into our minds through our eyes and our ears and our other senses and we just see things, perceive things, hear things as they really are. But that's I think a very misleading view of actually how perception works.” 2:17
“Just as color seems to be revealing the future of the world, the self seems to be revealing some persistent unchanging essence of me.” 0:24
Jürgen Schmidhuber · 2026-07-23
The deep-learning pioneer credits the compute supply chain, not the algorithms alone, for the current AI wave -- the same argument for backing hardware.
“A single human can't build artificial intelligences, and for that you need an entire civilization.”
“You also need the people who build computers... The hardware!”
“And it needs all these video gamers, who in the first place encourage companies like Nvidia to manufacture such fast processors because they need them for video games.”
What labs shipped 1
NVIDIA Robotics Research · 2026-08-05
It moves robot foundation model inference off the robot and onto networked server GPUs, changing the unit economics and BOM of fleet deployments.
“Existing systems are largely shaped by a single-robot, single-model assumption: inference is treated as an edge-computing problem handled by an on-robot or dedicated nearby GPU, and the serving objective is to minimize the latency of a single action model.”
“First, ROSA adopts shared GPU-pool serving, allowing a fleet of robots to access powerful server-class GPUs over the network in order to improve inference performance, battery duration, and GPU utilization.”
“Third, ROSA uses factory-objective-driven scheduling to maximize SLO-qualified factory productivity rather than minimizing individual request latency.”
What got funded 4
The Robot Report · 2026-08-07
A concrete bet that remote-operated humanoid fleets can bootstrap the real-world data needed to train robot autonomy.
“The path to full robotic autonomy will require both better hardware and massive amounts of real-world data.”
“I believe that model can help Avatar build one of the largest humanoid robot fleets in the world.”
The Humanoid Hub · 2026-08-06
Adds to a stacking queue of Chinese humanoid makers reaching public markets, a sign of investor appetite ahead of proven revenue.
“Shenzhen startup X Square Robot Technology, founded in 2023, confidentially filed for a Hong Kong IPO around August 4–6, 2026, with Huatai Securities and Morgan Stanley as joint sponsors.”
“X Square Robot last closed a funding round in June 2026 at a valuation of about 20 billion yuan (roughly $2.9 billion).”
“Its best-known product is Wall-B, a humanoid targeting China's household market, built on the World Unified Model (WUM) architecture.”
The Humanoid Hub · 2026-08-06
First humanoid maker to list on China's A-share market, and DeepSeek's tie-in signals convergence between China's top LLM lab and robotics hardware.
“Unitree Robotics set the final IPO price for its listing on Shanghai's STAR Market on August 6, 2026: 150.80 yuan per share.”
“According to the exchange filing, DeepSeek is investing 140.8 million yuan (about $20.8 million), acquiring 933,399 shares — 2.31% of the strategic placement tranche.”
“Unitree and DeepSeek, both headquartered in Hangzhou, will jointly develop AI models for humanoid robots.”
“The 150.80 yuan issue price implies a price-earnings ratio of 219.23 — 5.7 times the industry average of 38.56.”
The Humanoid Hub · 2026-08-04
The first hard financials from the humanoid market leader show margin and growth compression as it buys volume with a 70% price cut — the reference numbers every humanoid business plan now gets checked against.
“Revenue growth has slowed from 335 percent to around 40 percent, adjusted profit fell more than 52 percent in the first quarter.”
“The average selling price of humanoid robots at Unitree has fallen from around 590,000 yuan to 166,400 yuan — a decline of more than 70 percent.”
“Unitree, by its own account, is betting on a strategic pivot: away from the quadruped market, toward humanoid robots with AI-based capabilities.”
“The humanoid segment now accounts for more than 50 percent of total revenue — a structural shift that only began in 2024.”
“At the same time, Citic Securities forecasts a post-IPO valuation of 50 to 56 billion yuan — around 7.4 billion US dollars.”
Worth adding? 24
Mined from this week's material; not yet tracked. Add to config/people.toml or sources.toml.
Ancient Mysteries channel
Ran a substantive interview with Ilya Sutskever on what he's building next, directly relevant to the world-models and foundation-model beat.
Curiosity ⇔ Entangled channel
Hosted Karl Friston discussing active inference and the free energy principle, a direct match to that research interest.
EO channel
Featured 1X CEO Bernt Børnich on the $125M raise for physical AI/humanoid robots, squarely in the robot foundation model space.
Logically Answered channel
Profiled Jensen Huang's rise at NVIDIA, touching directly on the compute and semiconductors interest area.
MTS channel
Hosted Guillaume Verdon discussing why classical computing can't scale AI, a direct hit on the fund's thermodynamic computing interest.
Startup Fundraising channel
Covered Yann LeCun launching a new $100M AI fund, relevant given LeCun's centrality to world-model research.
Adeel Razi person
A recurring, senior coauthor with Anil Seth on Granger-causality and hemodynamic modeling work grounded directly in active inference and the free energy principle.
Colin Webb person
Co-founder alongside Anagbogu of the same freshly-funded humanoid robotics startup, a strong hardware-founder candidate despite no public profile.
Jack Hidary person
Coauthor with Guillaume Verdon on quantum Hamiltonian/variational thermalizer methods and CEO of a quantum-AI compute company, directly on-thesis for thermodynamic computing.
Jiajun Wu person
Stanford PI leading robot foundation model and world model research, appearing as a senior recurring collaborator with Fei-Fei Li across robotics/vision papers.
Ladi Delano person
Named founder of Moove, which just raised $250M to build robotaxi depot and charging infrastructure — a hardware-adjacent mobility play the fund's founder lens should weigh heavily.
Lancelot Da Costa person
Leads his own active inference and free energy principle research program, recurring senior collaborator with Anil Seth on neurophenomenology.
Lujie Yang person
Recurring Tedrake-lab collaborator on optimal control for robot locomotion and manipulation, directly relevant to robot foundation-model-adjacent control work.
Manling Li person
Appears as lead/recurring coauthor across 5 papers with Fei-Fei Li on embodied spatial intelligence and world-model planning, indicating she runs her own line of embodied-AI research rather than being one of many hands.
Mark Solms person
Collaborates with Karl Friston on free-energy-principle accounts of consciousness and affect, directly on-thesis for the active inference interest.
Max Simchowitz person
CMU PI working directly on world models and video diffusion for control, senior collaborator with Russ Tedrake on History-Guided Video Diffusion.
Nenye Anagbogu person
Co-founder of Avatar Robotics, a teleoperated-humanoid startup that just closed a $6.5M seed, squarely in the fund's European hardware-robotics founder focus.
Oriol Vinyals person
VP of Research at DeepMind and a central figure in frontier ML research, appearing alongside Yann LeCun's fund coverage this week.
Shubham Tulsiani person
CMU faculty running his own robot learning lab, senior collaborator with Abhinav Gupta on generalizable robot manipulation from human video (Gen2Act).
Simon Willison person
The only working AI practitioner in this list rather than a company or show name, and his hands-on evaluations of each frontier model release are a useful ML-research signal — though his beat is applied LLM tooling, not robotics or embodied AI.
Thomas Cohn person
Recurring collaborator with Russ Tedrake publishing on robot inverse kinematics and motion optimization, core robot-control research.
Tim Verbelen person
European researcher leading his own active inference work, recurring collaborator with Karl Friston on active inference and artificial reasoning.
Victor Dhédin person
Recurring collaborator with Majid Khadiv on dynamic-locomotion contact planning, a European robotics lab doing core embodied-AI control research.
Yejin Choi person
A senior, independent voice coauthoring the ESI-Bench embodied spatial intelligence benchmark with Fei-Fei Li, squarely in the fund's robot foundation model interest.