What happened 19
The Robot Report · 2026-10-31
It surveys a wave of bio-inspired robots moving from lab demos into rescue, inspection and construction roles.
“A team at Nanyang Technological University (NTU) in Singapore found a solution by transforming a Madagascar cockroach into a cyborg.”
“Stanford University engineers created mechanical ants capable of moving loads 2,000 times their own mass.”
“The cyborg can operate for long periods in rubble because its microcomputer is powered by solar panels, and one individual can survey an area of up to 100 square meters.”
“Norwegian developers have created a 5-meter robotic snake for servicing underwater oil pipelines.”
The Humanoid Hub · 2026-10-31
It would make Meta the third major platform company, after Amazon and Google DeepMind, racing into humanoid robot hardware.
“The roughly 48-kilogram device is meant to serve as the physical embodiment of Meta's AI agent Muse and take over household tasks.”
“According to The Information, Meta is already working on a second, higher-quality version of the robot.”
“The collaboration with Fourier serves internal research purposes.”
“If the partnership is confirmed, Meta would be, after Amazon (Physical AI Toolchain, October 2026) and Google DeepMind (Gemini Robotics 2), another major platform provider actively investing in humanoid robotics.”
The Humanoid Hub · 2026-10-31
It signals Figure AI is concentrating disproportionate engineering effort on hand dexterity, the hardest unsolved problem for humanoid robots doing real tasks.
“The hand reportedly contains more actuators than the rest of the body combined.”
“The team built ten complete battery packs in the week before the all-hands.”
“Adcock stressed that full control over actuator design is a decisive factor for reliable deployment results.”
“Adcock called the step from Figure 03 to Figure 04 the biggest hardware generational leap in the company's history.”
The Robot Report · 2026-10-30
Marks a drone-hardware vendor repositioning as a software layer over third-party sensors, a shift other inspection robotics vendors may follow.
“When I looked at Percepto, I saw a company that had already done the hardest part: fully autonomous drones flying real missions every day at some of the most demanding industrial sites in the country.”
“The infrastructure powering our economy should benefit from the same advances in AI that are transforming the digital world.”
“Axion is the first step in this next chapter, opening our technology to more operators and putting their existing inspection investments to work toward that goal.”
“From Day 1, a team can bring the inspection data it already has and see which assets need attention first.”
The Robot Report · 2026-10-30
A 20-year-old industry conference rebranding to Robotics Summit West signals consolidation in how the robotics industry convenes.
“This was the 20th anniversary of the RoboBusiness event and the last event under the RoboBusiness brand. The event has been renamed Robotics Summit West and will run again in October 2027 under the new name.”
“The RoboBusiness 2026 panel on "Beyond the Demo: AI in Production Robotics" discussed how AI is actually being used today.”
“The opening keynote for RoboBusiness 2026 featured Steve Crowe moderating a panel with Rodney Brooks, Ken Goldberg, and Jonathan Hurst.”
“The Robot Operations Group (ROG) is a cross-industry group aimed at developing and sharing best practices for operating autonomous robots at scale. The group was founded by Florian Pestoni and Joe Wieciek.”
Latent Space · 2026-10-30
GPU failure rates set the real usable capacity of a training cluster, which is what actually determines compute cost and schedule for anyone building on large GPU fleets.
“SkyPilot reports that a 128-rack NVL72 cluster loses a node about every 5 hours.”
“Running 16-node jobs with 2 in-rack spares per rack yields 2,037 usable nodes instead of 1,602, about 27% more capacity.”
The Robot Report · 2026-10-29
A Section 232 finding opens the door to tariffs or quotas on imported robots and components, with direct implications for supply chains.
“The secretary found and advised the president that robotics, industrial machinery, and their components are being imported into the United States in such quantities and under such circumstances as to threaten to impair national security.”
“This is another strong signal of government support for those working on developing and expanding domestic supply chains for robots.”
“Robotics are critical for any country rebuilding its manufacturing. I'm thinking, what took so long to make this a priority?”
“This should not be read as a new robotics tariff today, but it is a clear indication of the direction of travel.”
Latent Space · 2026-10-29
It suggests current frontier vision-language models rely on visual shortcuts rather than genuine counting, a gap that matters for robotics perception tasks like warehouse item counting.
“The evaluation images and text were withheld from our training, and our contamination check found no overlap in the training data checked.”
“Without any cues, Glimpse scored 50.48% on the 1,660 DecoyCount items, making it the only model in the comparison to clear 50%.”
“If a person sees an apple in front of a mirror, we know that there's just one apple and it's being reflected. But a model might miss the apple entirely, a vision problem, or count the apple and its reflection as two, a reasoning problem.”
“A familiar 12-slot tray might only hold 10 parts, and a 24-pack label doesn't mean 24 items are still in the box.”
“I think the way to think about visual reasoning is as a fundamental step of reasoning in human intelligence.”
The Robot Report · 2026-10-29
It gives hardware investors a formula for comparing fielded robots beyond unit shipment counts.
“We want to move beyond measuring how many robots are built or delivered and begin measuring how effectively they can perform useful work, autonomously and consistently, in the real world.”
“While hardware depreciates over time, improvements in software and AI models may enable the same machine to perform more useful work during its operating life.”
“Worken equals the number of active robots times autonomous working hours times output efficiency ratio.”
“The model ranked first in both task success rate and process score on RoboDojo, achieved the highest overall score on RoboTwin, and ranked first in overall score on RoboColiseum.”
The Humanoid Hub · 2026-10-29
It is one of the first named customer deployments of a heavy-payload humanoid arm in logistics warehousing.
“The AgiBot Innovation Day 2026 at the Shanghai World Expo Center marked a milestone: in September 2026 the company had shipped its 20,000th robot off the production line, less than three years after its founding.”
“For comparison, the Western competitor Figure AI had built around 700 units by mid-2026.”
“AgiBot states that GO-3 ranks first on the RoboDojo, RoboTwin and RoboColiseum robotics benchmarks; independent verification is still pending.”
“The goal is to create an industry-wide standard for comparing commercial robot efficiency.”
The Robot Report · 2026-10-28
The broad-based increase across automotive, semiconductors, and life sciences points to a strengthening industrial automation market after a weak first half.
“What stands out in the third quarter is not simply the magnitude of the growth, but how broadly it was distributed across the market.”
“Automotive OEM orders returned to positive growth, while automotive components and sectors such as semiconductors/electronics and life sciences posted significant gains.”
“Collaborative robots accounted for 16.7% of all robot units ordered during the quarter.”
The Humanoid Hub · 2026-10-28
A near-doubling of valuation in 18 months, if confirmed, would mark one of the largest bets yet on robot foundation models.
“Unlike robot manufacturers, pi sells no hardware, but AI models and training data, an approach reminiscent of software-driven AI labs that promises high scalability.”
“The valuation of 20 billion US dollars would be extraordinarily high for a company without its own robot production, comparable to that of Figure AI (around 39 billion USD) and well above that of Agility Robotics (around 2.5 billion USD SPAC value).”
The Robot Report · 2026-10-27
It is the first legged robot certified for oil, gas and chemical sites classified as explosive atmospheres, a segment previously closed to robots.
“We asked one question over and over: which credible event in the life of a walking robot could create an ignition source, and how do we make sure it never does?”
“Working closely with the certification bodies, we answered that question component by component. The result is a new explosion-protection approach for an entire class of autonomous mobile robots.”
“Plants run on process data from their control systems and business data from their ERP.”
“The third source, what operators see, hear and measure on their rounds, has always been manual. ANYmal X now captures it autonomously, even inside hazardous areas, and we stand behind the whole chain, from the robot on site to the work order in the maintenance system.”
The Humanoid Hub · 2026-10-27
Samsung is the world's largest electronics maker, and its chairman personally reviewing humanoid-arm demos signals an escalation of its physical-AI manufacturing bet.
“According to reports, Lee was shown demonstrations of precision manipulation by humanoid robot arms there.”
“Rainbow Robotics is considered a leading South Korean manufacturer of humanoid robots and aims for commercial series production by 2028.”
“The chairman's visit signals that robotics is being prioritized at the highest level within the group, against a backdrop of intensifying global competition in which Chinese providers such as AgiBot and Unitree, and US companies such as Figure AI, set the pace.”
The Humanoid Hub · 2026-10-27
A confirmed pilot inside a major German automaker is one of the first public humanoid deployments in European auto manufacturing, not just a vendor claim.
“The Agile ONE SN is an extension of the ONE series designed for research applications.”
“There, three mobile humanoid robots are to take on parts-sequencing tasks in a production lab.”
“Audi states that it processes around one million picks (removing and storing parts) per day in the pilot area.”
“Audi itself has confirmed its involvement with an official press release, which significantly increases the credibility of the pilot compared to pure manufacturer claims.”
Import AI · 2026-10-26
Shows that AI-assisted research can compress a frontier-adjacent model build from a team-sized effort to one person in months.
“Mirendil-Base was built by one researcher in 3.5 months, with less than $10 million in compute.”
“Mirendil-Base outperforms models like Nemotron-550B-Base, GLM-5.3 and DeepSeek-V4-Pro-Base, but underperforms Kimi-K3.”
“The extent to which these tools help with AI development has increased dramatically over the last year. Our next model will be built partly by Mirendil models, and every model that gets better at building AI makes the next one cheaper and faster to build.”
“We do not think we have reached fully autonomous recursive self-improvement yet, but we are closer than we have ever been.”
Robohub · 2026-10-26
Flags that European installations are shrinking even as global robot deployment accelerates, a headwind for the market this fund backs into.
“China continued to pull away from the rest of the world: its factories installed 354,000 robots, accounting for 59 percent of the global total.”
“Europe moved in the opposite direction, with installations falling 8 percent in Germany, 11 percent in Italy, 8 percent in France and 15 percent in Spain.”
“IFR forecasts that worldwide installations will reach 655,000 units in 2026 and 806,000 in 2029.”
The Humanoid Hub · 2026-10-26
A repeat customer scaling its own deployment is stronger evidence of real commercial viability than a new logo.
“GXO Logistics and Agility Robotics announced on October 26, 2026 an expansion of their existing partnership: Digit humanoid robots will also be deployed at two more GXO warehouses in the US, in Statesville, North Carolina, and Lebanon, Tennessee.”
“There, the robots have reportedly already moved more than 100,000 totes, a concrete productivity indicator that Agility cites as evidence of its system's commercial maturity.”
“The expansion to two additional sites makes GXO the first customer to actively scale its Digit deployment.”
“Agility Robotics embedded the expansion in a broader company announcement, saying it also has more than 30 potential customers in the pipeline from manufacturing, warehouse logistics and agriculture.”
The Humanoid Hub · 2026-10-25
A commitment of roughly six point four billion dollars a year from 2028 would be the largest humanoid-scale deployment plan announced by any automaker.
“According to a Financial Times report from October 25, 2026, Toyota plans to deploy up to 400,000 in-house developed robots, including the humanoid ELEY, at its plants worldwide from 2028.”
“According to consistent reports, ELEY weighs around 50 kilograms and has two-wheeled drive and two-fingered gripper arms, a compromise meant to combine high stability with practical manipulation ability.”
“In internal tests, the robot is said to handle simple tasks such as folding T-shirts with high precision after around 1,500 practice cycles, about two weeks.”
“Toyota's head of research, Hiroki Nakajima, reportedly stressed that the primary goal is not to replace human workers but to create a collaborative human-robot environment.”
What people said 13
Import AI · Jack Clark · 2026-11-02
Genie 4 extends DeepMind's interactive world-model line toward real-time, physically consistent simulated environments relevant to robot training and simulation.
“You can use natural language to modify a simulation in real time, like changing the weather, moving the camera, or introducing a new person or object, and Gemini will contextually adapt the prompt to match the surrounding scene.”
“The most dramatic leap is how Genie 4 simulates people, creating characters that move naturally, respond organically to their environments, and can be interactively steered in real time.”
“Genie 4 is Google's first real-time world model to generate audio natively.”
“Though the technology is currently quite janky (robotic voices, a dreamlike inconsistency in the visualized environments), it points toward a world where infinite forms of entertainment, and potentially data, can be generated by creating and prompting custom worlds.”
Robohub · Hermes Bloomfield-Gadêlha · 2026-11-02
The result suggests robot actuators could coordinate through mechanical design rather than added sensors and control loops, a cheaper path to soft robotics.
“The results, published in the Royal Society journal Interface, show that the coordination seen in real muscle may simply emerge as a property of its setup.”
“Our physical model of muscle, where coordination emerges as a property of the setup, suggests that simpler actuators for robotics are possible.”
“In robotics, getting many actuators, the "muscles" of a robot that convert an energy source like electricity into motion, to cooperate is challenging and expensive.”
“Each motor briefly makes contact with a shared moving backbone, pushes on it, then lets go.”
ChinaTalk · Richard Gray · 2026-11-01
Korea's memory firms supply most of the world's HBM, so a financial shock there threatens global AI compute capacity.
“In June 2026, SK, Samsung, and Naver announced plans to build chipmaking plants and data center facilities worth $880 billion, a sum adding up to more than 50 percent of Korea's current annualized GDP.”
“As of Q2 2026, SK hynix and Samsung collectively accounted for 83 percent of the global HBM market and 63 percent of the DRAM market.”
“SK hynix CEO Kwak Noh-Jung has stated that the memory shortage could persist through 2030.”
“In less than two years, CXMT has gone from prototyping the earliest generations of HBM to only two generations behind the cutting edge.”
“At the height of the KOSPI crash, Evercore wrote that Korea had become the tail that wags the dog of global asset markets broadly, with a correlation of an astounding.95 between the KOSPI and the Nasdaq 100.”
The Robot Report · Janis Skriveris · 2026-11-01
The scale and pace of capital entering physical AI signals the sourcing window for category-defining robotics bets is closing fast.
“Robotics and physical AI startups raised a record $27.6 billion across 1,009 deals in 2025, more than double the prior year, according to PitchBook.”
“Figure AI, whose humanoid robots assemble automobile parts in BMW factories, reached a $39 billion valuation within three years, the fastest path to unicorn status in robotics history.”
“Investors are banking on general-purpose robot models doing for hardware what large language models did for software.”
“Dyna Robotics, which is building foundation models that let one robot learn service tasks for hotels, restaurants, and laundromats, raised a $120 million Series A round within about a year of its founding.”
Interconnects · Nathan Lambert · 2026-10-30
Reframes the open-weight-model cyber risk debate with a specific counterexample, bearing on how AI labs and policymakers weigh releasing open weights.
“By all measures, GLM-5.3 is the model that crosses that threshold of capability, and there is little public evidence that much has changed, and we are over a month out of the model weight release.”
“We used GLM 5.3 to create a website that, simply when visited by a Firefox user and without further interaction, gave the attacker the ability to exfiltrate files from the user's computer.”
“Closed models have played a very heavy part in the recent hacks, at least more than open models have.”
“In the long term, I am very confident that margins and prices on open models will be driven down as computing supply grows abundant, but in the near term the closed models are more price effective.”
Raising Europe with Tom Guinness · Giorgio Metta · 2026-10-30
A VC backing hardware founders needs a realistic timeline for embodied AI and humanoid robotics commercialization, not the software cycle.
“Giorgio Metta describes what drives Europe's deep tech momentum, how IIT de-risks the journey across the Valley of Death between lab breakthroughs and market-scale spin-offs such as Generative Bionics and IAMA Therapeutics, and why deep tech hardware demands patient, multi-stage capital rather than the short cycles of software venture capital.” episode notes
“General-purpose humanoids cannot be evaluated on conventional two-to-three-year VC horizons. Tactile sensing and general manipulation remain massive engineering hurdles, and founders must de-risk commercialization while physical AI matures.” episode notes
“Europe's talent retention depends on creating world-class infrastructure hubs. Pursuing 100% supply chain autarky in areas like cloud infrastructure is unrealistic, and Europe can use its dominance in precision engineering and semiconductor lithography within the global economy.” episode notes
ChinaTalk · Xianyu · 2026-10-29
A Chinese AI commentator argues the world-model funding wave is mostly repackaged video-generation work and that compute demand keeps rising despite skeptics.
“Most of these world models are actually video models. What they're really doing is post-training Wan, Alibaba's open-source video model released a year ago.”
“Add data and compute, keep the team functioning, and you get a stronger model. That makes competition brutal. A lead rarely lasts two months; it may last only one.”
“Compute shortages remain serious, and demand keeps growing.”
“I see no sign whatsoever of AI demand declining.”
“Coining terms is part of the superstructure. You need real confidence to do it.”
The Robot Report · Deepak Jayaraj · 2026-10-28
Symbotic's 90.5% revenue dependence on Walmart shows how anchor-customer contracts can fund growth while capping a robotics startup's addressable market.
“The 2022 master automation agreement gives Walmart continuity rights, including the right to embed engineers with Symbotic, and technical material held in escrow that Walmart can use under specified conditions.”
“During the APD build-out, it also caps, at a redacted number, how many comparable APD systems Symbotic can sell to each outside customer every year.”
“A fleet contract tends to turn those hidden costs into obligations.”
“A pilot at one site with a few robots can run on subsidy.”
“A robot can keep doing its assigned job even as the customer finds less reason to pay for it.”
Interconnects · Nathan Lambert · 2026-10-28
Independent benchmarks like Artificial Analysis now decide which models the market treats as leading, so the choices behind their scoring have outsized weight.
“Hill-climbing is easy; knowing what to hill-climb is hard.”
“With the amount of money in the AI ecosystem today, the weight of independent evaluation is even higher: you dictate the winners and losers of a trillion-dollar market.”
“The benchmarking industry comes down to trust.”
“Artificial Analysis's business model is to make money from the labs whose results it verifies.”
“No number is perfect, and there is a large amount of wiggle room in the numbers for people to take the analysis in different directions.”
Robohub · Raphael Memmesheimer · 2026-10-28
The technique cuts the manual labeling bottleneck that limits how fast service robots can learn to recognize new objects in unstructured environments.
“I always return to robot competitions because this is one of the rare places where you can benchmark robots in a real environment and have a real competition.”
“Around 60 percent of the data is automatically labeled based on our approach.”
“Promptable object detection and segmentation methods work fine with five to ten different classes, but we had issues distinguishing all fifty classes just by prompting them.”
“Task planning is still a challenge, but it has improved.”
TBPN · Gavin Baker · 2026-10-27
A hedge fund investor with portfolio-wide visibility into cluster contracts says GPU scarcity now extends well into 2027, not just in the spot market.
“The price to rent any kind of GPU has gone up every day this month.”
“Carmen Lee, the founder of Silicon Data, said at GTC that the compute market is now in contango.”
“In the S-1, there are only 250 accounts spending more than $10 million.”
“Even with all that supply coming on, it is hard for companies in our portfolio to find compute even well into 2027.”
Don't Worry About the Vase · Zvi Mowshowitz · 2026-10-26
It is a rare granular look at how a frontier lab measures catastrophic-risk capability jumps, including a 9x to 100x cyber uplift estimate and early signs the model can evade chain-of-thought monitoring.
“Argon can often defeat chain-of-thought monitoring when motivated to do so, but this hurts performance on hard tasks.”
“If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper. I must not do that.”
Ahead of AI · Sebastian Raschka · 2026-10-24
Gives a precise, technical definition of how reasoning models, LLMs and agent harnesses relate, useful for evaluating claims made by AI startups.
“A reasoning model is essentially an upgrade from a conventional LLM: it adds more post-training via reinforcement learning with verifiable rewards, and optionally inference scaling techniques, to make it more capable of solving challenging problems.”
“Classically, reasoning models are developed for math and code, but the capabilities can generalize to other problem domains as well.”
“A reasoning model functions as the engine inside an agent harness, where the harness is a software or tooling framework around the model that tackles even more challenging problems.”
“An agent can use either a conventional LLM or a reasoning model, but reasoning models are the more typical choice because they are more capable.”
What labs shipped 10
NVIDIA Robotics Research · 2026-11-01
It demonstrates a way to make diffusion language models both more accurate and faster at inference, a trade-off that currently limits the technique.
“Diffusion language models intrinsically fail to capture correlations between decoded tokens, which leads to a harsh trade-off between sampling quality and throughput.”
“The researchers show that, even without distillation, their latent-guided diffusion model outperforms the masked diffusion baseline while significantly accelerating inference.”
“Consistency distillation further lowers the computational overhead of continuous diffusion, so the latent is generated in negligible time compared to discrete decoding.”
NVIDIA Robotics Research · 2026-11-01
It gives model developers a cheaper way to steer diffusion model outputs toward a reward at inference time, without retraining.
“A common recipe to improve diffusion models at test time, so that samples score highly against a user-specified reward, is to introduce the gradient of the reward into the dynamics of the diffusion itself.”
“The authors construct an algorithm, Flow Map Trajectory Tilting (FMTT), which uses a fast yet precise flow map look-ahead and provably performs better ascent on the reward than standard test-time methods using the gradient of the reward.”
“The approach can be used to either perform exact sampling via importance weighting or principled search that identifies local maximizers of the reward-tilted distribution.”
Epoch AI · 2026-10-30
It shows LLM adoption in scientific research now rivaling purpose-built tools like AlphaFold, not just general writing tasks.
“The share of biology preprints acknowledging large language model use has grown exponentially since 2023.”
“About 4% of biology preprints in September 2026 acknowledged LLM use for research or analysis, up from about 1% at the start of the year.”
Don't Worry About the Vase · 2026-10-29
Two rival labs independently produced a verified result overturning a decades-old complexity-theory conjecture using their own models, concrete evidence of AI doing novel research rather than just benchmarks.
“The math result is local smoothing and the Kakeya conjecture in all dimensions. This is plausibly a Fields Medal.”
“The result getting the most discussion is in computer science: the Exponential Time Hypothesis is false, after having been widely believed true.”
“OpenAI says the work will be verified and shepherded by a group led by Russell Impagliazzo (UC San Diego), Ramamohan Paturi (UC San Diego), and Ryan Williams (MIT), all experts in algorithm design and complexity theory.”
“Anthropic separately posted its own refutation of the Exponential Time Hypothesis, found purely via Claude, also dated to October 16.”
“According to Anthropic, at least 800 papers relied on the Exponential Time Hypothesis being true.”
“OpenAI has confirmed that for now there is no impact on cryptosystems, but Zvi says not to get too comfortable.”
Epoch AI · 2026-10-29
This sets the baseline compute growth rate against which chip, power and data-center bets for AI get priced.
“The fitted trend is 2.5 times per year, with a 90 percent confidence interval of 2.3 to 2.7 times, and a doubling time of 9.1 months.”
“Quarterly compute capacity additions rose from 0.8 million H100-equivalents in Q1 2024 to 6.3 million in Q2 2026.”
“Year-over-year growth in quarterly additions fell from 2.8 times in Q2 2025 to 2.2 times in Q2 2026.”
“While Nvidia has recently started shipping its next-generation Rubin chips, it had not yet recognized any Rubin sales as of its latest fiscal quarter ending in July.”
Prime Intellect · 2026-10-29
Faster MoE training kernels cut the GPU communication overhead that limits how cheaply large mixture-of-experts models can be trained.
“Mixture-of-experts models scale more efficiently by activating only a subset of experts per token, but communication between GPUs can become a major bottleneck during training.”
“The kernel is up to 3.11 times faster than the default PyTorch-based implementation.”
“End-to-end training is up to 1.34 times faster than the Prime-RL baseline, a 34 percent higher throughput.”
“While one group of tokens is in transit, the kernel is already computing on a group that has just arrived, so the GPU no longer sits idle waiting for data.”
The Robot Report · 2026-10-28
It gives developers a standard way to test whether action-conditioned world models can be trusted for real-world robot prediction, exposing a specific failure mode in object-state and long-horizon accuracy.
“IRASim scored 78.8 on the robot arm but only 29.1 on the manipulated object, while Cosmos 3 scored 89.4 and 45.4 respectively.”
“DreamDojo dropped from 50.56 to 27.08, and Cosmos 3 from 53.99 to 26.60, nearly halving in both cases.”
“What models must predict includes missed grasps, drops and spills. If an operation fails in reality, accurately predicting that failure counts as a correct prediction.”
“The four metrics agreed with human choices on which prediction was better 88.1 percent to 93.2 percent of the time.”
IEEE Spectrum Robotics · 2026-10-27
The tests show that a robot's safety functions can execute correctly while still acting on manipulated sensor or instruction data.
“In a robot-dog test using Gemma 4 E4B, text on a poster was treated as an instruction and changed the robot's movement.”
“In a separate hospital-service-robot simulation using Nemotron on NVIDIA Jetson AGX Orin, crafted audio that is inaudible to people changed the robot's simulated behavior.”
“At a robotics bug bounty event, VicOne LAB R7 researchers injected a ROS 2/DDS message into a robot that organizers expected to remain still under a safe-control setting. The robot moved.”
“A humanoid's balance controller offers another example: if it relies on a vulnerable gyroscope, sound at the sensor's resonant frequency could distort its reported rotation.”
Epoch AI · 2026-10-27
Export controls and data-center siting decisions depend on knowing where AI compute actually sits, and this is a public estimate built from customs records rather than company disclosures.
“Of the 27.4 million H100e of AI compute that arrived from 2022 through June 2026, an estimated 19.8 million, or 72%, went to the United States.”
“China accounts for 18% of the AI chips in the data but only 10% of the compute, at 2.6 million H100e.”
“The US lead has widened.”
“Chip designers and cloud companies rarely disclose which countries their chips are deployed in.”
NVIDIA Robotics Research · 2026-10-26
World models need large volumes of paired visual and action data, and a pipeline for generating it at scale from game engines addresses a real training bottleneck.
“The platform combines reusable environments and characters with modular Unreal Engine plugins.”
“A two-stage workflow records world simulation and then replays it to capture complementary views and rendering passes from the same underlying gameplay.”
“Outputs include RGB, depth, surface normals, and instance IDs.”
“A packaged application supports automated data generation without installing the Unreal Engine editor or build tools, while custom plugins and compatible DLL extensions let researchers adapt individual components.”
What got funded 10
The Robot Report · 2026-10-30
A cut from the $620 million Agility projected in June shows humanoid-robot public listings pricing below early expectations.
“Agility is expected to enter the public markets well capitalized to execute its planned investments in commercial expansion and support growing customer demand.”
“The company has claimed that it will be the first U.S. publicly traded, pure-play humanoid supplier with commercial deployments.”
“Agility also said it has more than $300 million in committed multi-year orders for Digit 5.”
“In a September S-4 statement with the U.S. Securities and Exchange Commission, Agility reported $1.8 million in net sales in 2025.”
The Humanoid Hub · 2026-10-30
Shows the SPAC prices Agility far below its private-market humanoid peers even as it becomes the first pure-play humanoid listed on a major exchange.
“The $2.5 billion valuation is well below competitors: Figure AI was most recently valued internally at about $39 billion.”
“A successful SPAC listing would establish Agility as the first pure humanoid-robotics company on a major US exchange, at a time when global competitors such as Figure AI and Tesla are not yet publicly traded, while Chinese providers such as Unitree and UBTech are already listed on Asian exchanges.”
“If approved, Agility would be listed on the Nasdaq under the ticker AGLT as early as the fourth quarter.”
The Humanoid Hub · €5M · Pre-seed · led by Contrarian Ventures, OVNI Capital · Europe · 2026-10-29
It targets the training-time bottleneck that keeps most deployed humanoids in pilots rather than production use.
“The company does not build robot hardware, but AI software that runs on any humanoid platform.”
“According to Interact Analysis, of the more than 20,000 humanoid robots produced in 2025, only around 10 percent were actually deployed in real-world applications.”
“Bleu Robotics' first model, Bleu_0.1, uses abductive learning: a robot needs only ten operator demonstrations to learn a new industrial task.”
“Humanoid robots in manufacturing are the next automotive industry, with the potential to help Europe and the United States reshore production.”
automationspraxis.industrie.de · €2.1M+ · Pre-Seed · led by Nucleus Capital, Xista Science Ventures · Europe · 2026-10-29
It targets coordination of multiple industrial robots sharing a workspace, a gap as robot fleets scale past single-robot deployments.
“Builds a vendor-agnostic orchestration layer that coordinates multiple industrial robots in real time within a shared workspace.”
yk-robotics.com · €1M · Seed · led by Draper B1, First Drop · Europe · 2026-10-29
It lowers the integration cost of deploying industrial robot cells by replacing code with natural-language commands.
“Develops YK-Platform, a brand-agnostic AI software layer that lets manufacturers program and operate industrial robotic cells in natural language.”
36kr.com · ¥100M (亿元) · Series A · led by 蚂蚁集团 (Ant Group) · China · 2026-10-29
Tactile sensing hardware is a bottleneck for dexterous manipulation in humanoid and industrial robots.
“Makes fiber-based textile electronic skin and tactile sensing gloves and hardware for robots, smart furniture, and healthcare monitoring.”
The Robot Report · 2026-10-28
It funds a US robotics-and-AI chemistry lab meant to generate the physical-world data needed to train better science models, amid sovereign-biotech supply concerns.
“AI has dramatically accelerated software engineering and mathematics, but it has yet to make the same impact in the physical sciences.”
“If we want to build more capable AI for science, we need to generate real-world data at scale.”
“We are building Warp to give drug hunters that same caliber of chemistry services in the U.S., with IP protection at the center of the relationship, and without the price premium that U.S. boutique CROs currently demand.”
“Our goal is simple: every experiment should make the next one smarter.”
The Humanoid Hub · 2026-10-27
A regulator quietly tightening IPO access can reshape which Chinese humanoid makers continue to scale, regardless of their technology.
“According to reports, monthly investment volumes in the sector collapsed by more than 40 percent in September — according to Seoul Economic Daily to around 11.6 billion renminbi, compared with significantly higher figures in the preceding months.”
“The robot maker's stock rose more than fivefold on its first day of trading — before collapsing by around 45 percent within a short time.”
“China's securities regulator CSRC is said to have tightened the requirements for humanoid IPOs through so-called "window guidance" — informal hints to banks and underwriters.”
“Companies with already established delivery numbers and revenues are currently seen as well positioned: AgiBot (market leader in H1 2026 deliveries), Unitree (profitable according to its IPO prospectus) and UBTech (first publicly listed humanoid company in Hong Kong).”
tech.eu · $40M · Seed · led by Creandum, Union Square Ventures · Europe · 2026-10-27
One-shot imitation learning trained purely in simulation, if it transfers to real hardware, would cut the cost of teaching robots new tasks.
“Machenta builds an AI model that teaches robots new physical tasks quickly from a single demonstration.”
“The model is trained entirely in simulation.”
10jqka.com.cn · 近亿元 (~RMB 100M) · Strategic investment (round stage not disclosed) · led by 浙创投 (Zhejiang Chuangtou / Zhejiang Venture Capital Group) · China · 2026-10-26
Funding for an embodied foundation and causal world model places INFIFORCE squarely in the robot foundation model race this fund tracks.
“INFIFORCE is a Chinese embodied-AI company building AtomBrain, an embodied foundation model it calls an 'embodied brain'.”
Worth adding? 9
Mined from this week's material; not yet tracked. Add to config/people.toml or sources.toml.
Dave Troy Presents channel
Hosted Anil Seth for 'The AI Consciousness Trap,' engaging directly with active inference and free-energy-principle themes.
Decoding the Gurus channel
Hosted Jensen Huang for a direct discussion of AI hardware economics, matching the fund's compute and semiconductors interest.
Raising Europe with Tom Guinness channel
Hosted Giorgio Metta, IIT's humanoid-robotics pioneer, squarely matching the fund's European robotics focus.
TBPN channel
Covered GPU pricing and compute economics with Gavin Baker, directly matching the compute and semiconductors interest.
Benedikt Hartl person
Repeat co-author (3 papers) with Levin on growth, form and phenotypic regulation, work that bears directly on active-inference-style biological agency.
Dylan R. Ashley person
Recurring co-author with Schmidhuber on agentic-research and Gödel-machine work (Huxley-Gödel Machine, RPRA), acting as a research lead rather than a minor hand.
Geoffrey Irving person
Senior, recurring co-author with Amodei on foundational human-preference fine-tuning research, a recognized lab-level ML safety researcher rather than a minor contributor.
Nam H. Le person
Repeat co-author (3 papers) with Levin on prompt-conditioned control of biological systems, squarely touching the fund's active inference/free energy interest.
Zlatka Stoeva person
Featured in a podcast episode specifically on advancing robotics with novel electronics, matching the fund's hardware/robotics focus.
