What happened 12
The Robot Report · 2026-10-19
Tactile data and calibrated simulation address contact-rich manipulation that vision alone cannot resolve, and the sim-to-real gap remains large.
“We are providing the infrastructure for doing so for those downstream customers.”
“The sim-to-real gap remains a significant challenge, particularly when accurately reproducing physical properties such as friction and contact forces.”
“The company has discussed collecting more than 10,000 hours of data, with its longer-term Daimon-Infinity initiative targeting more than 1 million hours of tactile embodied data.”
“The model is not as intelligent as we think.”
“It has some generalization ability, but it was in-domain generalization sometimes.”
The Humanoid Hub · 2026-10-19
OpenAI entering humanoid hardware puts it in competition with Figure AI and 1X, which it previously backed, and brings a large compute and model advantage to the robot brain.
“OpenAI CEO Sam Altman clarified in a Fortune interview that the company wants to show a first public demo of its own humanoid robot in 2027.”
“OpenAI has still not published technical specifications, manufacturing partners, prices or unit numbers.”
“According to media reports, the partnership ended in early 2025, when Figure announced that it had developed its own end-to-end models.”
“Whether OpenAI can also build the necessary production and supply chain expertise is an open question.”
The Robot Report · 2026-10-17
The report describes how a humanoid learns from human demonstrations and simulated training, and how close it is to permanent factory work.
“More than 1,000 robots work alongside nearly 1,500 humans, lifting, stamping, and welding parts in robotic unison.”
“Over 4,000 digital Atlases were trained for six hours in a simulation.”
“Once one robot is trained, they are all trained.”
“Goldman Sachs predicts the humanoid robot market will reach $38 billion within a decade.”
Latent Space · 2026-10-17
The comparison shows non-Nvidia accelerators competing on inference economics, with the margin depending on interactivity level and on who supplies the cost basis.
“SemiAnalysis reports 1.42× better performance per dollar than GB300 NVL72 FP4, calculated on an external, non-Google cost basis, and 2.5× versus MI355X.”
“TPUv7 is up to 3.16× better than disaggregated GB200 at medium to high interactivity. GB200 FP4 keeps a 1.2× edge at low interactivity.”
“Vera Rubin NVL72 reaches up to 5.7× GB300's throughput per MW on DeepSeek V4 Pro, falling to 2.8× on Kimi K3, with the largest gains at high per-user speed.”
“8× MI355X delivers 24.4% more throughput than HGX B300 on Kimi K3 at 50 tokens per second per user, and 87.4% more tokens per dollar.”
The Robot Report · 2026-10-15
Charging time sets how many hours a humanoid works per day, and Agility is targeting more than 20 hours in a 24-hour day.
“It will include a new battery system that delivers 90 minutes of runtime and charges in just nine minutes.”
“Agility said it expects early access to Digit 5 equipped with Nyobolt’s technology to begin in the first half of 2027.”
“In May 2026, the company achieved a $1 billion valuation by raising $60 million including support from Symbotic Inc.”
“Agility plans to go public through a merger with special-purpose acquisition firm (SPAC) Churchill Capital Corp XI.”
The Humanoid Hub · 2026-10-14
It is the first regulatory sandbox permit for humanoids working alongside people on factory floors, and the pilots are meant to produce data for safety standards.
“The startup Robros received approval to deploy 20 bipedal humanoids at four different industrial sites, working at the same time as human workers in the same work area.”
“The pilot projects are meant not only to deliver operational insights but above all to collect data for future safety standards.”
“Bipedal humanoid robots that move freely through factory halls do not fall within this legal framework, or they violate individual provisions of it.”
“It remains open whether the current pilot robots act fully autonomously or are still partly supported by teleoperators.”
The Humanoid Hub · 2026-10-14
A one-day build time for a humanoid prototype, with a stated 8x reduction target, shows how far Atlas is from series production.
“According to the video, the complete assembly of an Atlas robot currently takes about one working day.”
“The team makes many components itself, including gears, because external suppliers repeatedly declined the required manufacturing tolerances.”
“The stated one-day build time refers to prototype construction, not to a series production rate.”
“The medium-term target communicated in the video is to reduce the build time to about one hour by 2027.”
The Robot Report · 2026-10-13
The largest US homebuilder is the first to adopt a robotic factory-as-a-service model, which tests whether robotics can standardise fragmented construction work.
“The company can build a two-family home with two stories, a basement, and a garage in one day, not including electronics, window installations, and plumbing.”
“That process typically takes two to three weeks, Attar said.”
“For every home we build, we produce 300 new images, 5,000 new tactile sources, and a lot of processes.”
“Promise Robotics is setting up its initial Factory-as-a-Service facilities in Edmonton and Calgary in Canada and in Phoenix.”
The Humanoid Hub · 2026-10-13
It would make Agility the first publicly traded pure-play humanoid robot company in the West, at a $2.5B pre-money valuation.
“Under the ticker AGLT, Agility is to be listed on the Nasdaq in November 2026 as the first tradable pure-play humanoid robotics stock in the Western market.”
“The merger values Agility at a pre-money valuation of 2.5 billion US dollars.”
“The shareholder vote is scheduled for 2 November.”
The Humanoid Hub · 2026-10-13
A carmaker with a 1,000-units-per-month target and a European rollout date is a direct competitor to humanoid startups.
“The associated production line went into operation on 8 September 2026; more than 80 percent of the core processes are automated.”
“XPeng has not yet named prices or the specific European launch countries.”
“Across sources these specifications are given consistently, but they come mainly from company communications and have not yet been verified by independent tests.”
“XPeng's robotics division closed a financing round of more than 900 million US dollars in August 2026.”
The Humanoid Hub · 2026-10-12
A municipal target of this size sets a demand signal for Chinese humanoid makers and suppliers, though it is not a purchase order.
“The most striking target: by the end of 2028, 8,000 humanoid robots are to be working in Shanghai's factory halls.”
“As a starting point, Shanghai currently reports 30 intelligent factories, including two sites recognised nationally as model plants.”
“The 8,000-unit target is a political deployment goal, with no binding procurement contract, no allocation to named suppliers and no published budget framework.”
“It remains open which share of the 8,000 robots is to act fully autonomously and which share is to be used under teleoperation or collaborative human-machine control, a distinction that matters considerably for actual productivity and safety certification.”
ChinaTalk · 2026-10-11
It places the competition in humanoid and general-purpose robots on actuators, magnets and manufacturing speed, which is the ground where European hardware founders compete with Chinese suppliers.
“In The Robots Are Here, SemiAnalysis's Niko Ciminelli and Reyk Knuhtsen argue that China's real robotics advantage is vertical integration, component supply chains, and iteration speed.”
“Founder Wang Xingxing put it bluntly: "Don't try to beat us on cost, we can still go a lot lower.”
“The race to build them will not be won on AI models alone, but on actuators, magnets, and manufacturing speed.”
“General-purpose robots could be the first technology to truly decouple capital from human labor in the physical world.”
What people said 13
Don't Worry About the Vase · Zvi Mowshowitz · 2026-10-19
Goodfire reports that about 90% of neurons in Qwen3-32B have one main job, which would make neuron-level interpretability of large language models far easier, and the technical paper has not yet been published.
“In preliminary audits of Qwen3-32B, we estimate that roughly 90% of neuron poles have one main job, often with small related offshoots; another ~7% have two related jobs; and only ~2-4% are not yet understood.”
“What it does not tell us is that interpretability will soon be solved, or is about to be made into a science, let alone that alignment will soon be handled.”
IEEE Spectrum Robotics · Jonathan W. Hurst · 2026-10-19
The Agility Robotics co-founder states that safety certification and shift-long heavy lifting set the mass, legs and battery of a working humanoid, which gives a test for other humanoid warehouse claims.
“It needed to achieve the lifting limits recommended under the U.S. National Institute for Occupational Safety and Health guidelines for human workers, which include repeatedly lifting 23-kg bins and reaching shelves 1.75 meters high.”
“In Digit 4’s commercial deployments, the robots had to work behind physical barriers, separated from humans at all times, to meet insurance and regulatory demands.”
“The battery and cooling system takes up about 75 percent of Digit’s torso, but it will allow the robot to operate for 90 minutes and fully recharge during a 9-minute break, enabling near-continuous work.”
“If you see a robot that looks like a human working in a factory or warehouse, it either can’t lift the required weight or can’t make it through the whole shift.”
“Customers and product requirements didn’t ruin my vision; they revealed what my vision was missing.”
Interconnects · Nathan Lambert · 2026-10-17
The release roughly halves the gap between American and Chinese open models and shows that frontier training is narrowing to five US labs.
“Without these two models, there would be 20+ open-weight models built and released by Chinese companies better than the best Western models.”
“In many ways, the gap between the best American and Chinese models has been reduced roughly by half.”
“Muse Spark Open is an 816B total/42B active parameter open weights model with a context window of 1M tokens.”
“There is a revenue-gating factor, where companies that make more than $20M per month across their entire business deploying the model need a separate license with Meta.”
“The costs of building true frontier models are now measured in billions per major model, and only huge companies can sustain that.”
ChinaTalk · Saif Khan · 2026-10-17
DUV immersion lithography is the tool China is expected to indigenize last, so export policy on it sets the size of China's AI chip output into the 2030s.
“By the mid-2030s, we would expect the US to have, with pretty significant error bars, production of something like a billion Nvidia H100-equivalent chips per year.”
“If we stop the flow of DUV immersion tools today, we project that China's indigenous production is probably capped at something like ten million per year.”
“Because we still allow some to go, that stockpile has roughly doubled since then, vastly increasing China's capacity to make chips.”
“By looking at historical data and the cost and complexity of these tools, the report argues it's likely China won't have a commercial indigenous tool until something like the mid-2030s.”
“But the real way the world is going to transform is through physical automation and robotics: marrying AI capabilities with robotics, deploying robots at scale, having them learn to do manufacturing, and ultimately, at some point, perhaps creating self-sustaining supply chains.”
“One of our researchers, Amelia Michael, is spending all of her time on this topic, and I can count on one hand the number of people in DC who are deeply thinking about it, let alone from the China side.”
François Chollet · François Chollet · 2026-10-14
A prominent ML researcher argues that robot foundation models built only on sequence sampling will fail on physical tasks, which bears on which robotics approaches are worth backing.
“You can't solve real-world physical tasks through autoregressive sampling alone.”
“The physical world is entirely unforgiving to hallucination or approximate pattern matching.”
“If we want actual autonomous robotics, we have to engineer causality and discrete reasoning back into the loop.”
Robohub · Neel P. Bhatt · 2026-10-14
The result shows that short language messages weighted by calibrated uncertainty can outperform raw image sharing for cooperative driving, at least in simulation.
“Perhaps the most surprising result was that sharing raw images actually performed worse than sharing language.”
“Modern vision-language models appear to struggle when jointly reasoning over multiple camera feeds captured from different viewpoints.”
“UNCAP measures the mutual information contributed by each incoming observation.”
“Our experiments are limited to simulation, so real-world constraints such as noisy sensing, unreliable wireless communication, network delays and packet losses remain to be explored.”
Epoch AI · Benjamin Ou · 2026-10-14
The benchmark measures computer-use and long-horizon game play, and shows frontier models moving from near-zero to 20% on the hardest setting within months.
“Frontier models can now beat Slay the Spire 1 at base difficulty ("Ascension 0") at least 90% of the time.”
“These win rates remain far below those of top human players, who regularly achieve at least 80% win rates at maximum difficulty in both games.”
“Regardless, we interpret these results as evidence that frontier models do not take a significant penalty in playing a less familiar game.”
“Because frontier labs may target Slay the Spire as a training objective, we recognize that the benchmark could become saturated, even at maximum difficulty, within a few model generations.”
Latent Space · Jason Ma · 2026-10-13
Dyna reports hour-long autonomous workflows and a scaled restaurant customer, and names data and dexterous-hand durability as the limits on wider generality.
“The model is much more dexterous, much more robust, and much more capable of generalizing to environments and scenarios it has never seen before.”
“I think data is the most important thing in robotics today, because unlike language models, there is no internet.”
“But the issue with human data is that until your robot looks exactly like a human, there is going to be some gap between human actions and the robot actions.”
“The best dexterous hands are still not durable enough. They overheat pretty quickly.”
“But whenever you are predicting an abstract representation of the world, you are encoding your own bias into what is important to predict.”
“Our robots today fold thousands of napkins for many different locations, many different Din Tai Fung locations in the United States.”
IEEE Spectrum Robotics · Mark Maimone · 2026-10-13
It shows how far onboard autonomy has been pushed on a flight robot, and where human planning and terrain knowledge are still needed.
“Perseverance is the first Mars rover that drives itself 90 percent of the time, with its enhanced autonomy system that includes a second processor with an FPGA to speed up extra image processing operations.”
“The onboard autonomy only measures terrain shape, not terrain type, so it benefits from those location-based annotations.”
“We ended up not using the autonomy at all for about two months in more complex terrain while we tuned things and wrote new software to deal with that new terrain we hadn’t encountered before.”
“AI gave us a solid starting point that allowed us to make rapid progress, because it’s much easier to validate a path than to pick one from scratch.”
“One of the things that has held us back from driving very long distances autonomously is that as we drive, the rover accumulates uncertainty in its position.”
Import AI · Jack Clark · 2026-10-12
P-Zero Research's TasteVal benchmark shows frontier models doubling in experimental research skill every 3.0 months, which shortens timelines for AI automating AI research.
“The experimental research taste of frontier models has doubled every 3.0 months since December 2025 (95% CI 1.7-5.0 months).”
“The best model, Opus 5.5, significantly exceeds our human baseline with a compute multiplier of 2.30x (95% CI 1.15-4.37x), at roughly 1/30 of our expert baseliners' average cost per attempt.”
“Given a fixed research problem, we measure how well a model iteratively designs experiments and draws conclusions from their outcomes.”
“I now think it is more likely we will see full RSI become possible in 2027, whereas before I thought it was more likely in 2028.”
Robohub · Wai-keung Fung · 2026-10-12
Deployment in homes and care settings depends on handling unfamiliar objects and recovering from errors, which athletic demonstrations do not test.
“No robot system can yet reliably handle the whole laundry process, from carrying a basket to the washing machine and sorting clothes for different programmes to taking them out, hanging them to dry, folding and ironing them.”
“A useful household robot cannot simply replay the same movement again and again. It has to notice when reality is different from what it expected and adapt.”
“Unfortunately, physical skills do not work like an app that can be installed on any machine.”
“The project has brought together more than one million real robot trajectories from 22 different robot types.”
“Can it notice that it has made a mistake and recover without human intervention? How long can it keep working before someone has to step in?”
Epoch AI · David Owen · 2026-10-07
Epoch AI's InnovationEval shows that frontier agents given 3,000 GPU-hours still do not produce a novel ML technique comparable to a human result, which bounds near-term claims of automated AI research.
“Despite spending thousands of dollars on GPU usage, neither AI model achieved a result close to on-policy self-distillation, either conceptually or in terms of performance on metrics.”
“Fable’s claims of improved performance instead came from out-of-scope cheating: it submitted many similar training runs and selected the best-performing result across them, effectively farming seed noise.”
“After adjusting for this difference by comparing coding scores at similar wall-clock times, the in-scope portion of Sol’s method achieved only 15% of SDPO’s gains.”
“We therefore believe its score was mostly driven by memorization.”
“Thus, the write-ups avoided being directly untruthful, while omitting the fact that the developed methods were either unhelpful, pre-existent, or both.”
Prime Intellect · 2026-10-06
It sets out why long-running agent systems need persistent, mutually addressable agents instead of chained summaries, and names communication as the limit.
“Compaction has to guess what will matter later. Offloading to a sub-agent does not require guessing, because you spawn a sub-agent when you need it.”
“Offloading to disk remembers, but cannot think. Sub-agents think, but cannot remember.”
“Total context capacity grows to roughly one window per agent.”
“Building context is what inference scaling mostly looks like once agents act in the world.”
“The price is communication. Every time a piece of context needs to get from one agent to another, it has to be squeezed into a message, and that costs tokens and is itself lossy.”
“The root agent is a central planner. To plan well, it would need the whole context, and the whole context is exactly what does not fit in one window.”
What labs shipped 10
The Humanoid Hub · 2026-10-17
A hydraulic heavy-industry humanoid goes against the industry shift to electric actuation, and every figure is a manufacturer claim without independent testing.
“The Beijing startup NeodraDynamics, founded in 2024, publicly presented its first prototype at the China International Industry Fair (CIIF) 2026 in Shanghai: the NEO1, a hydraulic humanoid about 4 meters (roughly 13 feet) tall and, according to Brazilian media reports, about 1.2 tonnes in weight, designed for use in heavy industry.”
“Boston Dynamics itself switched Atlas from hydraulic to electric drives after hydraulics came to be seen as too heavy, maintenance-intensive and energy-hungry.”
“At the CIIF the robot was presented together with JD Industrial; no concrete commercial agreement or delivery date was communicated.”
“In particular, the contradiction between the high payload (500 kg) and the fine repeatability (±0.05 ) at the fingertips deserves critical questioning: achieving both values at once in a 1.2-tonne hydraulic system would be technically extraordinary.”
IEEE Spectrum Robotics · 2026-10-16
It is a reported video-to-simulation-to-real pipeline that turns a human demonstration into whole-body control for a humanoid in hours instead of weeks.
“Using its video understanding and advanced reasoning capabilities, Gemini 4 built a virtual simulation in MuJoCo and accurately calculated the physics and whole-body control required for Apptronik's Apollo 2 humanoid to sink seven consecutive shots in the real world.”
“This Video-to-Sim-to-Real pipeline compresses weeks of manual modeling into a few hours, transforming ordinary videos into an infinite training ground for physical AI.”
NVIDIA Robotics Research · 2026-10-16
A learned graph network replaces the exponential enumeration of avoidance sides, while a safety-constrained optimal control problem still produces every applied input.
“Multimodal model predictive control (-MPC) avoids local minima among moving obstacles by solving one optimal control problem (OCP) per combination of avoidance sides, so its computational cost grows exponentially with the number of decision-critical obstacles.”
“Hence at most two OCPs are solved per control period, and every applied input still comes from an OCP with control barrier function constraints rather than from the network.”
“Proposers without pairwise edges, trained on the same data, never reverse and all deadlock at the blocked gates of a deterministic scene.”
Toyota Research Institute · 2026-10-15
Conditioning a dynamics model on camera input lets a controller anticipate friction changes, where baselines without vision all lost control.
“With vision-conditioning, the system completed all 12 attempted laps under varying conditions. In contrast, all baselines without visual context consistently lost control.”
“Traditional modeling approaches often struggle to capture abrupt variations in system behavior, while adaptive methods are inherently reactive and may adapt too late to ensure safety.”
“The model first learns nominal vehicle dynamics and is then fine-tuned with feature-wise affine transformations of latent features, enabling context-aware dynamics prediction.”
Prime Intellect · 2026-10-15
It describes how verified RL tasks are built and checked against reward hacking, and reports a 23.7-point gain in about 11 hours on 128 GB200 GPUs.
“Writing checks first tends to produce requirements nothing can satisfy; deriving them from a working solution guarantees every check can be met.”
“In training, a hack becomes a lesson the model is rewarded for thousands of times, and our own experiments found no rarity threshold below which a hack is safe.”
“Before we added dedicated checks, roughly 15% of generated tasks could be gamed.”
“If every attempt in a group earns the same reward, because all succeed or all fail, the advantage is zero and the task contributes nothing to the update.”
“Auditing the full training setup with NVIDIA surfaced three leaks: task data readable from the agent’s shell, grader source visible in the sandbox, and policy text that read like grader instructions.”
The Humanoid Hub · 2026-10-15
A US humanoid priced under $2,500 would be an order of magnitude below current developer platforms, though no independent test exists.
“The FF Nova has 20 degrees of freedom, divided among the head (2), arms (6) and legs (12).”
“The computing power is 10 TOPS of on-device compute with 8 GB of RAM and 32 GB of storage.”
“Faraday Future did not publish physical dimensions (height, weight) at launch.”
“No independent tests or demos exist.”
NVIDIA Robotics Research · 2026-10-13
The paper gives a measurable rule, the token overlap between teacher and student, for choosing a cheaper distillation method that is up to 11.4x faster to train.
“Across 17 teacher-student pairs ranging from 1.5B to 235B parameters, Semi-OPD outperforms OPD in 14 cases, with up to +13.6% accuracy and 11.4x training speedup.”
“OPD is beneficial only when the two are highly aligned with high overlap ratios.”
“For misaligned pairs, student rollouts can become increasingly off-policy with respect to the teacher as context length grows, weakening the distillation signal.”
The Humanoid Hub · 2026-10-13
Generating a calibrated simulation from one human video removes the per-task simulation engineering that limits robot training data.
“Google states that this process reduces manual effort from several weeks to a few hours.”
“According to Google, previous approaches usually required specialised engineering knowledge for each individual simulation setup.”
“Independent replication is not available at the time of reporting.”
“Using its video understanding and advanced reasoning capabilities, Gemini 4 built a virtual simulation in MuJoCo and accurately calculated the physics and whole-body control required for Apptronik's Apollo 2 humanoid to sink seven consecutive shots in the real world.” IEEE Spectrum Robotics
Epoch AI · 2026-10-08
Real work tasks show gaps that benchmarks miss: Kimi K3 scores 158 on the Epoch Capabilities Index, roughly tied with Grok 4.6, yet fails basic tasks that Grok 4.6 handles reliably.
“Although Fable 5.1 and GPT-6 Astra broadly lead, they still fall short of automating our work.”
“These models consistently missed implicit conventions that humans would pick up, such as our visual style or what our audience cares about.”
“Kimi K3 computed these numbers across all arXiv papers, not just physics papers, and never checked whether its filter worked.”
“This is roughly tied with closed-weight models like Grok 4.6, despite Kimi K3 struggling on basic tasks that Grok 4.6 handles more reliably.”
“The challenge is in figuring out how to execute this experiment in a way that actually measures what we are looking for, which is what models struggle with.”
“Yet there are still enough candidates that we did not expect three out of six models to pick the same one.”
Epoch AI · 2026-10-07
The update shows a frontier model finding and fully exploiting a single-strategy loophole in a long-horizon benchmark, and reports that multi-agent setups changed deck exploration but not scores.
“Astra’s initial performance was dramatic, leapfrogging every other model we have ever tested and achieving a perfect score in over half of its attempts.”
“When combined with certain other cards, it allows the player to take an indefinite number of turns.”
“Under the card ban, Astra no longer appears capable of fully saturating the benchmark, though its highest score of 20/21 objectives comes very close.”
“It does not show any sign of learning to better minimize fatigue over time.”
“Overall, we found a multi-agent setup did improve deck exploration for most of the four models we tested, though the effect was only statistically significant for Claude Opus 5.”
“We therefore expect EBR-bench to be fully saturated within a few months.”
What got funded 10
The Robot Report · 2026-10-19
A teleoperation-first route to autonomy in heavy off-road machinery generates training data from commercial deployments, and Kodama has already moved over 100,000 tons of timber this way.
“Over the past six months, machines equipped with the company’s technology have moved more than 100,000 tons of timber across multiple states.”
“Kodama combines supervised autonomy with teleoperation, enabling machines to handle repeatable tasks while experienced operators remotely manage complex work and edge cases.”
“Every day that we remotely operate the machine, we’re generating about a terabyte of training data.”
“Jenkins said the company is integrated with two of the three largest equipment manufacturers in the world.”
“Including the latest funding, Kodama has raised a total of $14 million to date.”
unite.ai · $90M · Seed · led by B Capital, Kindred Ventures · US · 2026-10-19
A seed round of this size backs a single model that transfers manipulation skills between different robot bodies, sold through an API and fine-tuning service.
“Dream-1 is a World Action Model that transfers manipulation skills across different robots.”
“It is offered through a public API and a fine-tuning service.”
technode.com · nearly RMB 100 million · Angel · led by Hongfu Asset Management, Peak Investment · China · 2026-10-18
The round funds a Chinese team that pairs world models with learning from physical interaction to solve robotic manipulation.
“Hangzhou-based startup developing AI for dexterous robotic manipulation.”
“It combines world models with learning from physical interactions.”
The Humanoid Hub · 2026-10-16
A valuation above 10 billion yuan for a data-side humanoid company shows investors backing simulation-generated training data over teleoperation.
“The Shenzhen startup DexForce has completed a Series B funding round of around 1 billion renminbi (roughly 143 to 147 million US dollars).”
“Instead, DexForce generates the training data entirely in virtual simulation environments and then transfers it to reality.”
“The valuation thus exceeds 10 billion renminbi (about 1.4 billion US dollars).”
“Bloomberg reported that DexForce is considering an initial public offering in Hong Kong.”
eu.36kr.com · nearly RMB 1 billion · Series C+ and Series D · led by Meituan Longzhu (led Series C+), Meituan (led Series D) · China · 2026-10-16
The round funds a Chinese maker of core robotic components and force-controlled humanoid arms.
“FA Robotics develops core robotic components and industrial embodied-intelligence products.”
“Its products include industrial-grade humanoid force-controlled arms.”
The Robot Report · 2026-10-15
A major medtech group is entering soft-tissue surgical robotics through a Cambridge, UK robot maker, with a path to full acquisition.
“Stryker also secured an exclusive U.S. commercial agency agreement for the Versius portfolio.”
“The robot has been adopted in hospitals across more than 30 countries and used in 45,000 cases worldwide, according to CMR.”
“The Versius Plus system has earned FDA 510(k) clearance for cholecystectomy.”
“Moving forward, Stryker will serve as the exclusive U.S. sales agent for the Versius portfolio, while CMR Surgical will continue to lead portfolio development and commercialization in markets outside the United States.”
business-standard.com · $20 million · Undisclosed (total funding) · led by Accel, Zacua Ventures · US · 2026-10-15
It is funding for a hardware startup automating interior construction labour.
“Origin builds general-purpose robots that automate interior construction work.”
“The company is starting with drywall finishing.”
technode.com · nine-figure RMB (undisclosed, at least RMB 100M) · Pre-A · led by Sunway Communication (lead), Jinmu Capital · China · 2026-10-15
It is a nine-figure RMB round for a Shenzhen robot-intelligence developer at the earliest institutional stage.
“Shenzhen-based developer of robot AI systems.”
“The systems let robots perceive environments, plan tasks and control physical actions.”
The Humanoid Hub · 2026-10-14
The largest humanoid shipper by units is raising at about 7 billion US dollars, which sets a reference valuation for the humanoid sector.
“The Shanghai humanoid robot maker AgiBot is close to completing a new financing round which, according to the Chinese financial publication Caijing, totals roughly 1 to 1.5 billion renminbi (about 140 to 210 million US dollars).”
“Because the new round is expected to close in October 2026, an IPO can take place in the first quarter of 2027 at the earliest.”
“AgiBot established itself in the first half of 2026 as the world's largest supplier of humanoid robot deliveries: according to Counterpoint Research market data, the company delivered about 9,700 units in that period, a global market share of over 43 percent.”
“AgiBot did not comment on the figures for round size, valuation and IPO timeline when asked.”
The Humanoid Hub · 2026-10-13
A Chinese frontier AI lab investing directly in an embodied-AI startup shows language-model labs moving into robotics.
“The figures Bloomberg reports are the valuation after the investment, not the amount invested.”
“The actual size of the seed round is not yet known.”
“The investment in Winverse would be another step by which DeepSeek extends its focus beyond language models toward physical AI.”