What happened 20
The Robot Report · 2026-10-25
A large cobot vendor is distributing a European startup's robot foundation model through its integrator network, and the model already runs returns handling at Zalando sites.
“Together, the companies have already shipped hundreds of UR robots in dual-arm systems to live customer sites across Europe.”
“One brain serves every robot, with no retraining for each new item or task. And what one robot learns, every robot benefits from.”
“Cortex is trained in NVIDIA Isaac Sim, an open simulation framework, using NVIDIA Cosmos open world models.”
“In deployment, Cortex runs on NVIDIA Jetson Thor at the cell where inference must happen at line speed, and on NVIDIA DGX Spark for site-level development, validation and model updates.”
“The robots now autonomously process returned fashion items. This is the first phase of a broader initiative to scale returns automation across Europe.”
Latent Space · 2026-10-24
Artificial Analysis says the gap between the leading open-weight and proprietary models would narrow from 12 points to three once the weights are published, though serving a 2.8 trillion parameter model still needs substantial accelerator infrastructure.
“Kimi K3.1 Pro is a post-trained version of Kimi K3, with substantially larger post-training compute.”
“At launch, the API was available; the weights were not.”
“Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index, scoring 55, behind Opus 5.5 at 58 and Sonnet 5.5 and Muse Wonder Full at 56.”
“On the Intelligence Index, K3.1 Pro generated 55K output tokens per task, versus K3’s 48K, approximately 14% more, offsetting some of the token-price reduction.”
“Those comparisons concern per-token list prices, not measured cost per successful task.”
“Moonshot’s distinctive safety intervention was to remove cyber-related data from mid-training and reinforcement learning, rather than relying solely on inference-time refusals.”
The Robot Report · 2026-10-23
A European humanoid maker is building its own source of contact and force training data for world action models, and plans to use it to validate robots before customer deployment.
“Generative Bionics said the new facility will be operational by the fourth quarter of 2027.”
“The Genoa, Italy-based company plans to eventually open a second facility elsewhere in Europe and a third serving global markets.”
“This requires data that captures contact, force, motion, proximity, temperature, task constraints, and human interaction in realistic operating settings.”
“Each physical environment and robot configuration will also be recreated as a digital twin.”
Latent Space · 2026-10-23
The measurement gives the Ascend 950PR's compute and bandwidth against an H100, with 86% utilisation on compute-bound shapes.
“Measured at 600W, 325 BF16 TFLOPS (about 44% of an H100).”
“Roughly 1.4 TB/s effective bandwidth, but with 86% MFU on compute-bound shapes.”
The Humanoid Hub · 2026-10-23
South Korea is a leading producer of cars, ships and semiconductors, and the first government-approved humanoid deployment there gives Chinese suppliers an opening before domestic startups reach serial production.
“The South Korean government has approved a first official deployment programme for 20 mobile humanoid robots in semiconductor manufacturing, automotive, shipbuilding and logistics.”
“According to reports, Seoul proposes to raise spending on physical AI in 2027 to 3.1 trillion won (about 2.3 billion US dollars), an increase of more than threefold over the current year.”
“UBTech has reached a purchase agreement for its Humanoid U1 with the Seoul entertainment park Galaxy Robot Park.”
“Domestic Korean humanoid startups, however, are not yet ready for serial production; this explains why Chinese suppliers can initially fill a market gap.”
The Humanoid Hub · 2026-10-22
The delay shows that hand design, autonomy software and reliability still block the best-resourced humanoid programme from external sales.
“First deliveries to external customers are now to take place only towards the end of 2027, and not as previously indicated in mid-2027.”
“The AI models for autonomous work were not yet at the level Tesla requires for commercial use.”
“These units, however, are intended for internal testing and training purposes, not for sale.”
“Musk had originally promised 5,000 units for 2025; this target was not reached.”
ChinaTalk · 2026-10-21
The draft would put licences on new Chinese frontier models and on AI chip, memory and cluster design know-how, which changes the supply of open-weight models and compute hardware outside China.
“MOFCOM proposes to require licenses for the export of both weights and core private IP, including key checkpoints, underlying source code, architecture design, and technology for training optimization and inference acceleration.”
“As of October 20, 2026, no released Chinese model is estimated to have crossed this bar.”
“This restriction, and the following restriction on memory chip technology, does not restrict the chips themselves. Chinese companies remain free to sell AI chips and the memory used in data centers to U.S. companies.”
“Any technology exporter must register any exports of technologies among the above additions prior to export, including exports that have already begun but not yet completed.”
“By controlling VCT, China is hoping that the Korean companies will face the same yield struggles as CXMT.”
“This gives MOFCOM access to the technical specifications of every frontier Chinese AI model before it is released to anyone else.”
The Robot Report · 2026-10-21
Agility adds an agricultural-machinery maker to its customer list of Toyota, GXO and Schaeffler while it prepares a SPAC listing expected to close in the fourth quarter of 2026.
“The U.S. pilot, expected to begin in 2027, will support AGCO’s efforts to create more flexible, intelligent factories, beginning at its production site in Hesston, Kan.”
“Agility said its humanoid robot will initially support key assembly and production operations including sorting and stacking parts, as well as intelligent placement of totes for storage.”
“The business combination is expected to close in the fourth quarter of 2026, after which the combined company will operate as Agility Robotics and is expected to trade on Nasdaq under the ticker symbol “AGLT.”
The Humanoid Hub · 2026-10-22
NVIDIA becomes the only onboard compute supplier for Digit 5, and the announcement gives no unit numbers, volumes or productivity targets.
“On 22 October Agility announced that the general-availability version of Digit 5 will in future be based exclusively on NVIDIA IGX Thor as its onboard compute platform.”
“The early-access phase for Digit 5 is to begin in the first half of 2027, with increasing availability over the following six to nine months for manufacturing, warehouse and distribution operators.”
“Unit numbers, contract volume and exact productivity targets were not disclosed.”
The Information · 2026-10-21
The sell-off shows how the financing of GPU capacity for AI compute is exposed to interest rates and to delays in building data centers.
“A lot of the financing related to the data centers that are actually used by customers happens through something called a delayed draw term loan.” 33:58
“Where risks are introduced is in actually building and executing the facilities and getting the chips up and running to the point where they are producing revenue.” 35:17
“The unsecured corporate bonds are what you are seeing getting really hammered right now, and they are drawing all the attention.” 36:12
“You do have broader financing costs for AI-related projects raising across the board. You do have macro benchmark interest rates going up.” 36:12
“Insurers have already bought close to 800 million of one of the project-specific CoreWeave loans.” 39:27
“If the stock price is low, you have to sell more shares to raise the same amount of money.” 41:40
The Humanoid Hub · 2026-10-21
A Munich company with its own humanoid production in Germany is building on Nvidia's Jetson Thor and GR00T stack for university and corporate research labs.
“The ONE SN is designed as a research platform for universities and corporate research labs.”
“For 2027, Agile Robots plans 1,000 to 2,000 units.”
“Agile Robots is one of the few European companies with its own humanoid production on German soil.”
The Humanoid Hub · 2026-10-21
It is the first time Tesla separates Optimus production units from prototypes and pre-series units, although volumes are unconfirmed.
“The new Optimus units come off the line in an area that previously served electric vehicle production.”
“In its shareholder letter Tesla speaks of the first production-ready units the company has ever produced, which explicitly separates them from earlier prototypes and pre-series units.”
“There are no external checks of manufacturing quality and no independent acceptance tests.”
The Humanoid Hub · 2026-10-21
It is the first consumer-oriented humanoid from a Chinese carmaker's robotics brand, announced without price, delivery dates or independent testing.
“Mobo is 0.9 meters tall and has 22 degrees of freedom, which enable walking and other body movements.”
“AiMOGA states that it has so far delivered more than 3,000 units globally, to over 60 countries and regions.”
“An exact sales price, delivery dates after the pre-order phase or unit targets were not communicated at the time of the announcement.”
“The available information comes from company communication and is not yet supported by independent tests or deployment reports.”
The Humanoid Hub · 2026-10-21
It shows Unitree moving toward security and rescue applications, while autonomy and capability remain unverified.
“Whether the actions shown are performed autonomously or teleoperated or choreographed cannot be determined from the video material.”
“Unitree has not announced full specifications, price or market launch period.”
“Until independently verified, all capabilities shown are to be classed as manufacturer claims.”
The Humanoid Hub · 2026-10-21
It shows a humanoid taking over machine-tending tasks without rebuilding the production line, which is the stated advantage over conventional industrial robots.
“An AI task planner translates plain-text instructions into machining routines and coordinates the collaboration between the humanoid robot and the CNC machine.”
“According to Mitsubishi Electric, conventional industrial robots require costly line conversions for comparable tasks; humanoid robots are meant to avoid this conversion effort.”
“There are no independent tests of the system and no results from real production environments.”
The Robot Report · 2026-10-20
A defense customer expanded a robotic inspection contract fourfold within six months, which shows inspection robots moving from pilot to fleet-scale deployment.
“The news came just six months after an initial $71 million IDIQ contract award.”
“Gecko said its AI and robotic technology can save the U.S. Navy up to $8 million in repair costs per ship and return vessels to sea quicker.”
“Komodo scans through thick, non-skid deck coatings without stripping or damaging them, giving the Navy the ability to see the health of assets without taking it apart.”
“It produces a full-deck heatmap of corrosion and material thickness, enabling repair teams to pinpoint damage and plan work before a ship enters drydock, which was previously possible only after maintenance had begun.”
The Humanoid Hub · 2026-10-20
Korea is building actuator and motor capacity as a supplier to humanoid makers, in a market where Chinese firms ship 70 to 95 percent of units.
“The aim is to build a national supply chain for key components of humanoid robots, including actuators, electric motors and other mechanical assemblies.”
“South Korea is responding with this private initiative to the dominant market share of Chinese manufacturers; according to calculations, Chinese firms currently control 70 to 95 percent of global humanoid shipments, depending on the market researchers' methodology.”
“The focus on supplier components instead of end products is strategically notable: South Korea is positioning itself as a supplier to global robot makers, not as a direct competitor to companies such as AgiBot or Unitree.”
The Humanoid Hub · 2026-10-19
A continuous endurance run exposes hardware wear in a humanoid, though the figures come only from the company and no test protocol is published.
“The robot was connected to an external power supply by a cable, so it was not running on battery.”
“Figure has not published a detailed test protocol, for example on whether maintenance interruptions took place or whether a single unit was used throughout.”
“The test measures only walking time and distance on level ground, not performance under changing conditions, payload or in real work environments.”
The Robot Report · 2026-10-13
A state agency renewing a companion robot programme after three years shows sustained public-sector demand for care robots.
“NYSOFA and Intuition Robotics created a new three-year, $2.1 million agreement.”
“Participants engage with ElliQ an average of 48 times per day, up from 29 daily interactions in reporting from previous years.”
“More than 3,500 older adults applied to participate in 2025 alone.”
Latent Space · 2026-10-13
Meta Superintelligence Labs reached near-frontier agentic performance 14 months after its founding, at a lower cost per task than Anthropic's Fable 5.1 and Opus 5.5.
“Muse Wonder and Muse Wonder Full score 52 and 56 at max reasoning effort, each a material gain over Muse Spark 1.3 (max) at 48, with a release gap of only 6 weeks.”
“Wang says Meta continued scaling both pre-training and RL since Muse Spark shipped in April.”
“Output: both models use about 54k tokens per task, down roughly 10% from Spark 1.3’s 60k.”
“Wonder costs $4.52 per index task, driven by high input usage.”
“Wonder Full in Muse Code scores 68, tying Claude Code with Sonnet 5.5 (68) and ahead of Opus 5.5 (66).”
What people said 18
The Robotics Stack · Parsa Bakhtiari · 2026-10-25
A wheeled, modular platform with about 30 units sold tests whether robotics splits into separate hardware, model and deployment businesses.
“And we believe that robotics will be the same, where someone has to provide the hardware, someone will provide the brain, and someone will provide the deployment.” 1:08
“Anyone who has worked with legs in robotics realizes how difficult it is, because in robotics you have to go to the same point over and over again with great precision. And that's very difficult to achieve with legs. That's why we opted for a platform with wheels.” 10:41
“Actuator costs have come down so much that they are becoming a smaller and smaller fraction of the bill of materials, of the system cost. We obtain our actuators from a third-party supplier in China.” 20:14
“We measure success by the amount of maintenance and support requests we receive. So we know an implementation is successful if we don't hear from the customer.” 16:31
“I can say that we have sold around 30 units and we are building more.” 21:23
“I think we're reaching a point where you can do a great demonstration, but can you do it over and over again every day, day after day, with minimal human intervention? So I would take many of these demonstrations with a grain of salt.” 31:00
Interconnects · Florian Brand · 2026-10-25
Moonshot's K3.1 Pro is a roughly three trillion parameter open-weight model that ranks near the top of Artificial Analysis, and its release policy on cyber capability sets a precedent for how open labs handle dual-use risk.
“You want very broad access for everyone to find all the bugs and to patch all the bugs, which is net positive for the world.”
“Exploiting and being able to go through or wreck a whole organization is way harder, because you need very long-horizon capabilities, and you need to persist through all the stages that you penetrate.”
“In general, the Artificial Analysis benchmarks are rather short in terms of time horizon, whereas something like ProgramBench is really long and also very explorative.”
“To fit the model, you need B300s, one node per model replica, which is kind of awkward, versus other models which you can even fit on H100s, or H200s rather, which are more readily available.”
“The model creator has a few weeks of exclusive access before they release the weights.”
Simon Willison · Simon Willison · 2026-10-23
An open-weight model within three points of the top proprietary model changes what frontier capability costs to run and to self-host.
“Artificial Analysis scored Kimi at 55, ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53) and behind only Claude Opus 5.5 (58).”
“This is the first time a (soon-to-be) open weight model has scored this highly against the frontier.”
“The US dollar pricing is $2.00/million tokens for input and $10/million for output, the same as GPT-6.1 Sol and half the price of Claude Opus 5.5.”
Latent Space · Ankit Maloo · 2026-10-23
It gives a test for which open problems automated search will solve next: few conceptual leaps and a small branching factor.
“AI shrinks it far enough that the search pattern itself flips, and everything downstream follows from that.”
“With this checker, it is now rational to spend enormous compute on uncertain ideas. You find out whether a given idea works or not within minutes.”
“This predicts that models would do well where b^k x compute cost (where k is the number of hints) is small.”
“This may indicate one way that AI systems have an edge: it’s not just that they can try all known methods, but they can play for longer and in more treacherous waters than mathematicians without getting overwhelmed.”
“Even that took 10,000 agents, with a total of 300B tokens used in 88 hours.”
ChinaTalk · Aqib Zakaria and Hamish Low · 2026-10-23
If different silicon suits each inference phase, HBM-free accelerators become viable and the question of whether China can build them matters for compute supply.
“You need HBM for AI inference; inference decode is a memory-constrained operation, and only HBM was able to reach bandwidth speeds to serve models effectively.”
“But the market has matured since then, and there is no longer one architecture to rule them all. Different silicon works better for training, prefill, or decode.”
The Information · Seth Goldstein · 2026-10-23
Self-funded humanoid and compute buildout means Tesla does not need outside capital to scale Optimus production, which sets the pace competitors must match.
“The positive free cash flow shows Tesla is taking a more capital-efficient approach, which I think is the right call here.” 36:18
“Optimus is now in production, with plans to sell to commercial customers by the end of next year.” 36:43
“I think they might hold off on building the Texas plant for a year or two until the California plant really gets ramped up, and that is an efficient use of capex.” 38:25
“I think Tesla would rather be slightly capacity constrained than have a bunch of factories sitting idle, not running, and be a giant waste of free cash flow, as well as dragging down earnings from higher operating costs.” 38:44
“I do think that Tesla's ability to mass-manufacture robots at scale, and its complex production and logistics operations, should be a great training ground to allow robots to do a wide variety of tasks that make them useful to a customer.” 44:16
Latent Space · Erik Bernhardsson · 2026-10-22
Fixed multi-year GPU capacity leaves customers with idle hardware off-peak and shortages at peak, which is the case for pooled, elastic compute and for rethinking how capacity is sold.
“The biggest downside of GPU scarcity is not that it's hard to get GPUs, but that almost all capacity is sold as a three year rectangle starting three to six months in the future. That's not what demand looks like for most users.”
“A lot of workloads in real life do not look like flat, rectangular workloads. They scale up and down throughout the 24-hour cycle, they may grow over time, and they may have certain burst periods.”
“We can aggregate that capacity and get very smooth workloads, very smooth aggregate demand, which we can manage very efficiently.”
“We started seeing tremendous growth coming from a number of Neolabs doing things like reinforcement learning on Modal sandboxes. So sandboxes is now about a third of our revenue.”
“The faster you can get that wake-up, go-to-sleep cycle, the more you can create sandboxes that only run for a very brief period, and that will make compute much more efficient.”
“I do think 2027 might be the year of non-NVIDIA accelerators.”
Dwarkesh Patel · Pushkar Ranade · 2026-10-22
A 2x ceiling on leading-edge wafers is far below the 3x yearly growth in AI compute, so chip supply limits how fast model training can scale.
“Not 10X, but probably 2X, yes.” 50:38
“We have to make the most of the capacity that we already have installed.” 48:42
“We are locked into that one tool, one supplier, and we rely on that.” 49:04
“If your yield is already pretty high, getting every 1% extra yield is extremely hard because it is already a finely tuned process.” 50:01
“I would argue that the supply crunch is larger for memory today than it is for compute.” 59:46
“You can be off by a few percent, but you need to be on time.” 56:49
Don't Worry About the Vase · Zvi Mowshowitz · 2026-10-22
The report ties the size of the US-China AI gap to two variables policy can change, China's chip access and its distillation of US models, with half of the restrictable chips sitting in data centers abroad.
“The current 16x US training compute advantage would normally yield a 13-month lead, but we estimate adversarial distillation halves this to around 6.5 months.”
“Around half of the chips the US could restrict are in data centers outside China that the Chinese AI ecosystem will access remotely.”
“Create a new end-use control to restrict China's remote access to training clusters abroad.”
“They estimate China will only progress at half-speed in 2027 and end the year 13.1 months behind.”
“If America stopped releasing models, it would take 13 months for China to match our current best models.”
ChinaTalk · A. Chen · 2026-10-22
Usable footage costs five to seven times raw footage, so cleaning rather than filming sets the economics of humanoid training data.
“He places the cost of collecting one hour of raw data at around 100 yuan, and the cost of generating usable data at around 500-700 yuan per hour.”
“NVIDIA’s team attached cameras to the foreheads and wrists of over 1,000 people, which they used to film 20,854 hours of first-person footage: mostly everyday tasks, such as washing dishes and cutting vegetables.”
“Unless egocentric data collectors can harvest, say, 10 billion hours of data, we can’t outperform UMI in the market.”
“Egocentric data is not a versatile product. Footage is difficult, if not impossible, to resell: customers often have specific requirements for the lighting, background, and movements for every video, and customized data is often protected by exclusive IP.”
“According to Dr. Wang, startups raising over 1 billion RMB will usually pursue egocentric data used for pre-training, while those raising less than 1 billion RMB focus on post-training methods that require little to no egocentric data.”
“Chinese data collectors recruit subjects to film themselves at specialized centers, or from their own homes, instead of attaching cameras to workers in their workplaces.”
In Conversation With · Guillaume Verdon · 2026-10-21
The episode covers thermodynamic computing, which builds hardware that computes with the physics of noise and energy, and links it to Friston's active inference and the energy cost of computation.
“We sat down in the woods in SF to follow the thread from a kid obsessed with the cosmos to Google Quantum, Extropic, and learning to build intelligence from the physics of nature.” episode notes
“How do our beliefs shape the futures we build? How do we stay embodied as AI accelerates? What new aesthetics might help us imagine what comes next?” episode notes
Latent Space · Evan Conrad · 2026-10-20
Lenders finance GPU clusters against forward prices, and a market whose prices are easy to manipulate or detached from deliverable capacity raises the financing cost of every cluster.
“A market that’s very easy to manipulate and a market that has a lot of basis risk.”
“The reason why they can’t make it is because they would then not be independent enough for people to trade on it.”
“If you don’t trade on it, then you don’t have enough liquidity for lenders to finance against the forwards market.”
“You can make $40 an hour or so selling that as inference, any model basically. That’s if you use it at a hundred percent, and maybe if you only have like 50% utilization, so like $25 an hour.”
“To get closer to something that can be standardized, you first need to make all the operations of the thing the same.”
“You put the data center next to the power rather than next to the customer.”
ChinaTalk · Nick Brown · 2026-10-20
Lithography is the tool China is least able to build domestically, so DUV immersion exports decide how large the US lead in AI chip production stays through the 2030s.
“We estimated the entire installed base and the rate of imports: about ninety systems per year in 2024 and 2025, worth over ten billion dollars, close to two-thirds of ASML’s entire immersion DUV exports.”
“We estimate a single NXT:1980i can produce over a hundred thousand Ascend chips per year, even assuming China’s much worse yields.”
“Even if they diverted every system to their most advanced fabs and spent 100% of that capacity on AI chips, not a single smartphone, at the basic level of physics and lithography throughput, they’d be well below US production, something like a tenth of it.”
“If you cut off imports, you close the door on anything approaching US-China parity.”
“But some degradation in yield and throughput, starting maybe six to twelve months out and continuing over time, is reasonable to expect.”
Import AI · Jack Clark · 2026-10-19
The Anthropic co-founder argues the next phase of AI is robots and automated laboratories, which would raise demand for scientific equipment and robotics hardware.
“They call this the singularity. I believe that this year we have gained convincing evidence that it has now begun.”
“In robotics, AI systems are now increasingly able to autonomously operate laboratory equipment remotely, making it possible for us to further automate experimentation in domains from biology to materials science, making progress faster in areas like batteries and fuel-cell development.”
“When the exam launched in January 2025, top scoring AI systems obtained a score of 8%. As of this week, the highest scoring system is Claude Opus 5.5 with 67.7%.”
The Robot Report · Faisal Mahmood · 2026-10-18
Actuator performance depends on the motor, drive, feedback, gearbox and thermal path together, which is the integration problem hardware robotics startups have to solve.
“The actuator is a system. The motor, inverter, current sensors, rotor position feedback, gearbox, brake, output encoder, thermal path, communications interface, and higher level robot software all influence the result.”
“FOC does not automatically create a precision actuator. Rotor angle accuracy, current sense quality, motor parameter identification, sampling timing, inverter dead time, loop bandwidth, mechanical transmission, and tuning all affect the final result.”
“A motor side encoder can report the rotor position accurately while the joint output still contains transmission error.”
“A controller that handles acceleration correctly can still create an over-voltage problem during aggressive regenerative braking.”
“A controller advertised with a large peak current is therefore not automatically suitable for a motor that requires that current continuously.”
ChinaTalk · Richard Bian · 2026-10-16
A product lead at a Chinese open-weight lab argues that AI gross margins, not capital expenditure, decide whether the business model works, and describes how his lab uses small and domain-trained models.
“Every time the hype cycle repeats, capex swarms in, which becomes a self-fulfilling prophecy for what we build next.”
“For SaaS and mobile internet, the business model was viable because the gross margin on software is probably 95%.”
“Thirty or forty percent is already pretty good on inference.”
“Our new flash model is a 124-billion-parameter MoE with 5.1 billion active parameters.”
“Right now, people judge models in very similar and mostly superficial ways.”
“General intelligence is like university or high school. It makes you a generally capable person, ready for the world.”
Don't Worry About the Vase · Zvi Mowshowitz · 2026-10-15
Chips that can only run inference on an approved model would give regulators a hardware mechanism for halting frontier training, which bears directly on chip production and compute policy.
“David Krueger and others come out with a 210 page analysis on the feasibility of a hardwired pause of frontier AI training.”
“Chip production is switched into chips hardwired so they can only do inference on a particular approved model, and then existing training compute is rotated out over time.”
“The conclusion is that yes, this is technically feasible if the political will is there. The economic impacts would, they predict, be highly manageable.”
“That is very different from free, and I believe they are underestimating the true costs.”
Staer · Tobias Sjölin · 2026-10-08
A continuous measurement of empty driving and idle time gives a baseline for judging whether a warehouse needs more robots or better use of its existing trucks.
“At one live site, we measured fleet activity continuously over 25 days. Two numbers stood out: 40% of total driving distance was empty driving, and 29% of shift time was idle.”
“A WMS tracks transactions: pick confirmed, pallet moved, order shipped.”
“Empty driving is often structural: a truck finishes a drop-off far from its next pickup, or a layout forces a longer return path than the task needed.”
“A continuous empty-driving and idle-time baseline shows exactly where capacity is already being lost before anyone adds a robot to the mix.”
→ DeepSeek's open stack should lower token prices; estimates put its gross margin on tokens at up to 90% in What labs shipped
What labs shipped 9
IEEE Spectrum Robotics · 2026-10-23
A reduction of more than 1,000x in GPU-hours for humanoid motion tracking lowers the compute needed to train whole-body control.
“In three hours, on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for seven days on 128 GPUs.”
“It reduces GPU-hours by more than 1,000× and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline.”
“After 9.8 hours on two GPUs, SuperMimic achieves zero failures.”
Agile Robots · 2026-10-22
Moving planning and prediction to GPUs in the mobile network would remove onboard compute as the limit on a mobile robot's weight and battery life.
“Reflexes stay on the robot. Safety functions and motion control react instantly, without waiting for the network.”
“The heavy thinking, such as planning and predicting movements, moves to powerful GPUs nearby.”
“Our dual-arm Franka Research 3 setup draws on this computer to perform individual tasks in server assembly.”
“This is an early step in a lab setting, but it shows that the architecture can work.”
Agile Robots · 2026-10-22
Humanoids cannot work beside factory staff until their contact force and balance behaviour can be certified, and the ISO standards for this are being written now.
“When a robot arm stops, it is safe. When a humanoid freezes mid-stride with a load in its hands, it may topple over and create the very hazard the stop was supposed to avoid.”
“Depending on which part of the body it might touch, and in what situation, the robot adjusts its force and speed.”
“Because this logic sits inside the robot itself, companies that deploy humanoids would not have to define every contact limit from scratch for each new application.”
“The other is written specifically for robots with actively controlled stability, meaning robots that need continuous control to stay upright and could become unstable if they lose power.”
Franka Robotics · 2026-10-21
The post sets out a concrete agentic control design in which an orchestrator has no robot access, each specialist agent has its own tool server, and simulation tools cannot move a physical arm.
“The orchestrator has no direct access to the robot tools.”
“Each specialist accesses only the tools required for its role through a dedicated Model Context Protocol (MCP) server.”
“The simulation agent only receives tools for controlling the digital twin and cannot move a physical arm.”
Latent Space · 2026-10-21
The release gives first-party figures for throughput and revenue per accelerator-hour, but the cache-heavy workload and omitted costs limit any margin estimate drawn from them.
“It reports 155 trillion tokens per day from one cluster, 96% input-cache hits, 400 ms average time to first token, and 230 tokens/s generation.”
“The technical account says the KV-cache management and load-balancing components are not yet fully released, because they remain coupled to internal infrastructure.”
“The vLLM team reported decode throughput within 6% of DeepInfer’s published results on GB200 under a matched configuration.”
“The within-6% result is decode-stage parity, not reproduction of DeepSeek’s whole service.”
“Revenue is not profit, and an accelerator-hour is only one part of a production service’s cost base.”
Interconnects · Nathan Lambert · 2026-10-20
The piece estimates the margin DeepSeek earns on its own tokens and expects the published stack to lower serving costs for other inference providers.
“Estimates point to DeepSeek making up to 90% gross margin when selling tokens on this inference stack.”
“DeepSeek followed the above news by publicly releasing its internal inference stack, DeepInfer, which is likely far more efficient than existing open-source tools like vLLM and SGLang for its models.”
Robohub · 2026-10-19
A single layer of muscle cells now produces enough force to swim, which makes thin biohybrid actuators a lower-cell-count alternative to bulky 3D muscle.
“At its fastest, the robot can swim a distance of about four times its body length in one minute.”
“The new robot is the first example of a very thin, two-dimensional, muscle-powered robot capable of locomotion.”
“They found that cells grew in better alignment, and produced the most force, on stiffer gels.”
“The film was light enough that the cells were able to stick to the gel when they contracted, rather than peeling away.”
Toyota Research Institute · 2026-10-15
It shows that fusing inertial odometry with multi-camera perception improves lane estimation at racing speeds while meeting real-time rates on a vehicle.
“We introduce a new dataset for 3D lane detection in racing, featuring more than 250k images from multiple camera feeds and inertial measurements taken with a Lexus LC 500 driving on a closed circuit.”
“We propose modifications that permit frames to be processed at rates of almost 300 Hz while retaining high predictive performance in the racing application.”
“We show F1 scores above 0.9 and lateral MAEs below 0.18 m in vehicle deployments.”
Sereact · 2026-10-15
Improving a robot policy on real hardware from a few dozen attempts, with no simulator and no human correction, reduces the robot time and operator attention each new task needs.
“After 50 attempts per task, its success on three returns-handling tasks rose from 60-70% to 85-100%.”
“This is not straightforward, because you cannot tell whether an action was bad just by looking at it. The same grasp might lead to success or to failure far into the future.”
“On-policy methods need so much fresh data that they have mostly stayed in simulation.”
“FLEX turns a knob only when the new suggestion agrees with the direction that knob has been turning recently.”
“Stronger pretrained VLAs perform better but saturate at 60-70% on real-world tasks.”
What got funded 10
The Robot Report · $150M · Series A · led by Andreessen Horowitz (a16z), F-Prime · US · 2026-10-24
A $150M round for a simulation-trained manipulation model shows investor appetite for physics-based reinforcement learning over human demonstration data.
“Founded in 2024, Eka is building a vision-force-action (VFA) model to develop an intuition of physics through reinforcement learning at scale.”
“Robots using its AI can pick a wide variety of objects outside the training distribution at a speed of roughly 1,500 operations per hour.”
“Eka also said it believes that robots should master physics, not mimic human data, to be truly general-purpose.”
“In simulation, the company can generate thousands of hours of robot experience in minutes, varying objects, environments, and physical conditions.”
“Andreessen Horowitz (a16z) and F-Prime, a venture capital fund affiliated with FMR LLC, the parent company of Fidelity Investments, led the round.”
The Robot Report · $90 million · Seed · led by B Capital, Kindred Ventures · US · 2026-10-22
A video-trained model that controls unseen low-cost robots zero-shot, offered through a public API, tests whether a hardware-agnostic robot foundation model can work as a product.
“Dream Labs said its Dream-1 flagship model works across robots from different manufacturers.”
“Because it learns from video rather than from a single machine, its capabilities transfer across hardware, Dream Labs claimed.”
“Dream Labs supports robots today that cost between about $2,000 and $10,000.”
“Dream-1 currently handles tabletop tasks such as picking and placing objects and folding laundry, with support for wheeled mobile robots and dexterous hands to follow.”
The Humanoid Hub · $30M · Pre-seed · led by Redalpine, Notion Capital · Europe · 2026-10-22
A European startup replaces separate tactile sensors with force measurement inside the actuators, which would simplify robot hands if it works outside the lab.
“Instead of attaching separate touch sensors to the robot limbs, the company measures forces directly inside its own actuators.”
“At the same time, Enact has so far published no independently verified performance data, benchmarks or deployments.”
“With the new funds, Enact plans to increase its workforce fivefold by the end of 2027, accelerate hardware development and supply first industrial partners with prototypes.”
The Humanoid Hub · 2026-10-22
The maker of a roughly $1,400 consumer humanoid may become one of the first pure humanoid companies to list, which sets a public-market reference for the sector.
“According to a Bloomberg report of 22 October 2026, the Beijing startup Noetix Robotics (officially Songyan Power Beijing Technology) has confidentially submitted a draft prospectus to the Hong Kong Stock Exchange (HKEX).”
“In March 2026 Noetix closed a Series B of about 1 billion yuan (roughly 138 to 145 million US dollars).”
“According to the company (not confirmed by independent reports), Noetix is already producing several dozen units per day.”
tech.eu · over €28M · Undisclosed (growth round) · led by Índico Capital Partners (co-lead), Lince Capital (co-lead) · Europe · 2026-10-22
It is a European defence and infrastructure drone company funded at growth stage.
“Portuguese maker of AI-powered unmanned aerial systems and vehicles.”
“With a cloud management platform, for defence, security and critical infrastructure monitoring.”
techcrunch.com · $215 million · Undisclosed (new company launch round) · led by Lightspeed, IVP · US · 2026-10-22
A $215 million launch round shows investors funding robot automation of laboratory chemistry.
“Warp Sciences trains robots to synthesize and test drug candidate compounds.”
“Automating wet-lab work for drug discovery.”
technode.com · Not disclosed (total equity financing this year exceeded RMB 2.5 billion) · Pre-Series B-plus and Series B · led by Beijing Artificial Intelligence Industry Investment Fund, Beijing Robotics Industry Development Investment Fund · China · 2026-10-22
The amount shows how much capital Chinese investors are putting into dedicated processors for robots.
“Huixi Technology develops AI system-on-chip processors.”
“Computing platforms for embodied AI and robots.”
techcrunch.com · $115M ($90M equity and $25M debt) · Series D extension · led by Nimble Partners (equity lead), Silicon Valley Bank (debt) · US · 2026-10-20
Autonomy retrofitted to existing aircraft is drawing growth capital, including debt.
“Reliable Robotics develops autonomous flight systems that can be installed on existing aircraft.”
“It also develops an in-house solid-state collision-avoidance radar.”
dealstreetasia.com · nearly $400M · Series E · led by Boyu Capital · China · 2026-10-20
A round of this size shows capital concentrating in autonomous last-mile delivery vehicles in China.
“Neolix develops Robovans, autonomous logistics vehicles that deliver goods and cargo in office and industrial parks, campuses, residential areas and central business districts.”
“The vehicles are autonomous logistics vehicles that deliver goods and cargo in office and industrial parks.”
technode.com · tens of millions of RMB · Seed · led by Haowen Tianshu Robotics, Shanghai Lanyun Industrial Development · China · 2026-10-20
The seed round funds robots for vertical surfaces and fixtures, which are harder than floor cleaning.
“Jingji Intelligent builds cleaning robots for vertical surfaces, countertops and sanitary fixtures.”
“Jingji Intelligent builds cleaning robots for vertical surfaces.”
Worth adding? 5
Mined from this week's material; not yet tracked. Add to config/people.toml or sources.toml.
In Conversation With channel
Hosted Guillaume Verdon on thermodynamic computing and the future of AI, which matches the thermodynamic computing interest directly.
Guillaume Lajoie person
Leads his own neuro-inspired machine learning research and is a recurring coauthor with Blaise Agüera y Arcas on compositional structure.
M. Cerezo person
Leads a line of research on quantum machine learning and coauthored the Challenges and Opportunities in Quantum Machine Learning review with Guillaume Verdon, who works on thermodynamic computing.
Mark Sandler person
A senior Google researcher who recurs as a coauthor with Blaise Agüera y Arcas, most recently on meta-learning update rules.
Yanbo Zhang person
The most frequent coauthor with Michael Levin in this week's material (4 papers), working on learnable novelty and intelligence in basal cognition.
