NVIDIA Bought Hugging Face and Never Said Robots

NVIDIA's blog post announcing that it will buy Hugging Face for $12.93 billion runs through the platform numbers, the open-ecosystem pledges and a quote from Jensen Huang, and it never once says the word robotics. No LeRobot. No Isaac. No physical AI. I went looking because the thing NVIDIA agreed to buy is where almost every open robot foundation model in this field gets downloaded from, and NVIDIA already sells the chips those models train on. The trade press supplied the robotics angle. The buyer did not. Two smaller stories from the same week point the same direction. A marketplace opened three days earlier that grades and sells humanoid training data in Hugging Face's robot file format, and a lab claimed the fine-tuning step in between can be replaced by a video in a context window. Three layers of the open robot-learning stack moved toward being owned in eight days.

NVIDIA buys Hugging Face

NVIDIA agreed on 2 September to acquire Hugging Face for $12.93 billion, putting the distribution layer for open robot foundation models inside the company that sells the training hardware. Huang's commitments are unusually specific, including a pledge that NVIDIA compute will not be required. The objection worth taking seriously is about ranking. A hub decides what comes first, and that tilts a field without anybody acting in bad faith.

What a model hub is

If you have trained or fine-tuned anything in the last few years, you have pulled weights from Hugging Face. The word "hub" undersells it.

Model hub: the place trained model files live and get downloaded from. It hosts the weights, the configs, the dataset shards, the model cards, the download counters and the search index. It is also the default from_pretrained string in a million scripts, which makes it a habit as much as a service.

Hugging Face hosts roughly 3 million models, 500,000 datasets and 1 million applications for about 18 million developers. Almost none of it is robotics. The robotics part is a library called LeRobot.

The terms

Item	Figure
Total consideration	$12,930,300,000
To shareholders	$11.9B
Equity for employee retention	~$1B
Last private mark	$4.5B (2023), on ~$400M raised
Definitive agreement signed	2 September 2026
Announced	3 September 2026
Expected close	First half of 2027
Rank among NVIDIA acquisitions	Second largest, after the ~$20B Groq asset purchase in December

The close date is the line to hold onto. First half of 2027 puts the antitrust review ahead of the deal, so nothing here is final. Clément Delangue told CNBC that Hugging Face approached Huang over the summer, and that there were other bidders.

LeRobot, and why this is a robotics story

Robot data is scarce in a way image data never was. You can scrape the web for pictures. You cannot scrape it for a robot arm's joint angles under load, or the wrist forces during an insertion. Those have to be produced by somebody operating hardware, which is why the open robotics community leans on shared infrastructure far harder than the language modelling community ever had to.

LeRobot: Hugging Face's open robotics library. It defines a common dataset format for robot demonstrations, ships training and evaluation code, and hosts the policies. Its value is that a model somebody trained on their arm can be pulled down and run on yours, because data and weights arrive in a shape you already parse.

Since July 2026, LeRobot has carried NVIDIA's own Isaac GR00T 1.7 and Isaac Teleop, with Cosmos 3 announced as coming. That partnership predates the acquisition by two months, and it turns a software-platform deal into an embodied-AI story. Every open vision-language-action model this show has covered gets pulled from that host, which now belongs to the vendor whose accelerators those models were trained on.

What Huang promised

The pledges in the announcement are more concrete than the usual language about respecting the community.

Commitment	What it would take to break it
"Hugging Face will remain an open platform for the entire AI ecosystem"	Gating uploads, or tiering access by hardware target
Continued support for open-weight models	Deprioritising or deprecating open-weight hosting
Multi-cloud and multi-accelerator	Cloud-exclusive features, or CUDA-only tooling in the default path
NVIDIA compute will not be required to build on or deploy through the platform	Any default path that silently needs an NVIDIA GPU

That last one is the hard promise, because it is falsifiable, and companies do not usually write falsifiable sentences into acquisition posts unless they mean them. The pushback on the record is about whether a promise is the right instrument at all. Nithya Ruff of the Linux Foundation put it best in VentureBeat's roundup: "Neutrality is a discipline a company must choose time and again, not a promise it makes once."

Discoverability drift

The mechanism people worry about here is duller than censorship, and much more likely. A hub ranks. When you search it, something comes first, something loads fastest, something gets featured. If artifacts tuned for NVIDIA hardware quietly rank higher and download faster, the field tilts and no policy was ever written. Ranking is a set of weights somebody tunes for engagement, and engagement is downstream of what already runs well on the most common hardware.

There is an observable test, and this is the part I keep coming back to. Download-ranking parity for artifacts not tuned for NVIDIA hardware is measurable today, from outside, with the public API. Picking matched model pairs and tracking their rank and download velocity over time would answer this, and publishing that series is a weekend's work. Nobody is doing it, and the useful version starts before the deal closes so there is a baseline to compare against.

A second effect gets much less attention. Owning the chips and the hub together lets you see what the whole open community is training, months before any of it ships. No pledge addresses that, because it requires nobody to act badly.

Why a hub is harder to fork than a codebase

The reflex answer to platform capture is "fork it," which works for code because code is text. A hub is petabytes of weights, plus the download counts that establish which checkpoint is canonical, plus the incoming links from every paper and every README, plus the muscle memory of developers who type the same address without thinking. You can mirror the bytes. You cannot mirror the defaults. Duane O'Brien of the Open Source Initiative, in the same roundup, makes the prior point: "These platforms would not exist without Open Source software and open collaboration." That is the community's real leverage, and leverage is not the same thing as a default. Fork talk exists, ModelScope and community mirrors among it, and unless there is a real failure of neutrality it is expected to stay talk.

The word the announcement never uses

I fetched the post to check rather than take the trade press at its word, and the trade press is right. The words robotics, LeRobot, Isaac and physical AI never appear. The most consequential infrastructure event of the year for open robot foundation models was announced as a software-platform deal, and two readings fit. Either NVIDIA does not think the robot ecosystem is the point of a $12.93B bet, which is plausible when robotics is a rounding error against data-centre revenue. Or it does, and would rather nobody said so while an antitrust review is still ahead. I lean toward the first with real uncertainty, and I would rather flag the ambiguity than pretend a blog post can resolve it.

Kinetic Blocks, three days earlier

On 1 September, a small Oslo company called Kinetic Blocks opened a gated-beta marketplace for humanoid training data. You list a dataset, it gets graded, a buyer licenses it at checkout, and chain-of-custody documentation comes attached so rights are settled before training starts.

Element	Detail
Accepted categories	Egocentric human video, teleoperation recordings, robot execution data
Submission format	LeRobot v3.0 or the seller's proprietary format
Quality grade	KBQS, the Kinetic Blocks Quality System, applied to every dataset
Pricing	Sellers set prices, the platform takes a margin, commercial licence attaches at checkout
Access	No open signup, both sides approved one account at a time
People	CEO Lars-Fredrik Forberg (Oslo), board chair Erik Ålgård, data quality lead Leo Terry (Boston)

This show has spent four episodes on data quality, and this is the first attempt I have seen to turn quality into a price signal instead of an argument. The grade is what gives me pause. KBQS is the product, its rubric is proprietary, and so robot demonstrations now have a score and a checkout button with the scoring inspectable by one company in Oslo. Price a market by a number nobody can audit and data quality acquires a number without anyone establishing that it improved.

And note the format line. An independent business built its interchange layer on LeRobot v3.0, a file format whose maintainer, three days later, had a chip vendor behind it. Nobody planned that, and it is still the shape of the week.

Limitations

The deal has not closed and may not close on these terms. No ranking data exists yet in either direction, so discoverability drift is a mechanism with a plausible path and zero measurements, and I would rather say that plainly than dress a worry up as a finding. The Hacker News threads were rate-limited when I tried them, so the community read comes from secondary coverage. Kinetic Blocks is a gated beta with no published volumes and no prices.

Further reading: - NVIDIA's announcement, the terms and the four commitments, with no mention of robots - TechCrunch on the $12.9B confirmation, deal structure and Delangue on the process - VentureBeat's practitioner roundup, the on-record reactions and the discovery-neutrality question - Isaac GR00T 1.7 and Isaac Teleop on LeRobot, the July partnership that makes this a robotics story

Skild AI's S1 and the video prompt

Skild says you can specify a robot task with a single human video dropped into the context window, run the shipped weights, and skip fine-tuning entirely. The reported gap on unseen tasks is 66% step success against 9% for a language-prompted baseline. There is no paper behind any of it, and the most interesting claim in robot learning this month is a blog post with videos.

Skild announced S1 on 25 August, four days before this week's window opened.

Fine-tuning versus prompting

Getting a robot foundation model to do a new task normally means collecting demonstrations, running more training, and shipping new weights. That loop takes days at minimum, and a small team in practice.

In-context learning: giving a model the examples at inference time instead of training on them. The weights never change. The examples ride in the prompt, and the model works out the pattern from what it sees. This was a language-model discovery, and moving it to physical action is the whole claim here.

S1 puts a video of a person doing the task into the context alongside the live camera feed. Skild's phrasing is "no fine-tuning, no post-training." They call it omni-bodied, covering quadrupeds, humanoids and static arms, though the blog never says which platforms ran which task.

What Skild reports

Metric	Reported figure
In-distribution long-horizon tasks	96%
Unseen out-of-distribution tasks, step success	66%
Language-prompted VLA baseline, same tasks	9%
Pretraining scale for that result	100,000 hours
Value of one in-context demonstration	~380 post-training examples
Data quality spend	$3 on quality control per $1 on collection
Task length	Up to 10 minutes
Demonstrated unseen tasks	Potting a plant, cooking pancakes, pour-over coffee, kit assembly

Pretraining mixture is robot teleoperation, UMI hand-camera video, egocentric video and simulation.

Why a video specifies more than a sentence

"Make a pour-over coffee" is a terrible task specification. It says nothing about pour rate, nothing about the order of operations, nothing about how long you wait for the bloom. A video says all of that for free, at 30 frames a second, without anyone writing it down. Language is a lossy channel for physical work, and the loss lands in the continuous quantities that decide whether contact-rich tasks succeed.

Three dollars of QC per dollar of collection matches what I have seen running large-scale auto-labeling pipelines. The expensive part of a data flywheel is never the collection.

The callback to episode three

Episode three found three labs (tau-zero VLA, Zetta and EXIMO) independently putting the intelligence in an outer loop around a frozen policy, and left open whether that scaffolding is a permanent architectural layer or the thing the next generation of weights absorbs and deletes. S1 is the absorption case. The adaptation moves inside the context window, the weights stay frozen, and the scaffolding disappears. It landed four days after that episode aired. A serious point for absorbed, and faster than I expected.

Limitations

There is no paper. No architecture, no parameter count, no public API, no comparison against a named competitor. The 380x multiplier is Skild's arithmetic on Skild's tasks, and the 9% baseline is described but not identified. Every number above is self-reported and self-selected, and the video evidence is chosen too. I want this to be true. I cannot check any of it, and it is the situation the audit papers this show keeps covering were written to address.

Further reading: - Skild AI on S1, the task videos and the 380x claim - The Robot Report on the launch, independent write-up of the same announcement - ZimaBlue, the other answer to the same question, arguing that 120,000 hours of pretraining gets you generalisation without any of this

What happens when you speed the task up

Two policies scored an identical 100% on a dexterous insertion task at the speed the demonstrations were collected. Speed it up and the expert loses 16 points while the learned policy loses 34 and 48. Equal nominal success told you nothing, and every success rate in this article was measured at a single execution speed.

This paper got three upvotes on Hugging Face. Qwen-Drive got 370. I think it is the most useful negative result of the week.

The setup

The setup is a dexterous insertion task run two ways: a scripted expert, and an ACT policy learned from that expert by imitation. Then the authors varied the one axis nobody varies, which is how fast the task gets executed.

ACT (Action Chunking with Transformers): a standard imitation-learning architecture that predicts a short sequence of future actions at once rather than one step at a time. Chunking helps with compounding error, and it is the backbone of a great deal of published manipulation work.

The numbers

Condition	Scripted expert	ACT policy
Nominal speed (demonstration speed)	100%	100%
Sped-up conditions	Drops ~16 points	Drops 34 and 48 points
Maximum demonstrated speed	84%	53%

The failure analysis is where it gets sharp. 35 of ACT's 47 failures at maximum speed were insertion misalignments, so the policy was failing at the precise-timing step rather than at random. The authors' conclusion, in their words: "Equal nominal task success therefore does not imply preservation of expert performance across execution speeds."

Why this matters for every other number in this article

Nominal success rate is the field's universal metric, and it is a single-speed measurement that nobody labels as one. Facet-0's 82% is a nominal number. Skild's 66% is a nominal number. Every LIBERO figure in the papers this week is a nominal number. It makes them narrower than they read, and it means the phrase "matches the expert" has been doing work it cannot support.

I have watched this in perception too. A model matches the average precision of the one it replaced, then falls apart at 70 mph instead of 45, because the training distribution had a speed profile nobody wrote down.

Further reading: - Does Imitation Learning Preserve Temporal Robustness in Dexterous Manipulation?, the expert-learner comparison, the speed sweep and the failure breakdown - Peg-in-Bench, a fully 3D-printable modular insertion benchmark with configurable tolerances, which is the right tool for asking this question at scale

Facet-0 and the wrench you expect to feel

A group at NTU built a manipulation policy that predicts each action paired with the force reading it should produce, then scores candidates on whether the contact matched rather than on whether the task progressed. On five sub-millimetre computer-assembly tasks it reports 82% mean success against 15% for the strongest baseline. The mechanism is the interesting part, because it puts physical tolerance inside the loss function.

Sub-millimetre insertion

If you have pushed a stick of RAM into its slot or seated a ribbon connector, you know the shape of this problem. The clearance is a fraction of a millimetre, and vision stops helping the moment the parts touch, because the contact patch is hidden by the parts themselves. What you go on is how the part pushes back, and whether that matches what a correctly aligned part feels like.

Wrench: in robotics, the six-dimensional vector of forces and torques measured at a point, usually the wrist. Three force components, three torque components. It is the full description of how the world is pushing back.

How Facet-0 works

A conventional manipulation policy predicts what the joints should do next and gets graded on whether the task finished. Facet-0 changes both halves. It uses flow matching to generate action sequences, each carrying an expected wrench profile, so the prediction is "do this, and here is what it should feel like." An Action-Wrench Critic then scores the candidates on contact outcome rather than task progress. Training data is ManuFacet-1K, 1,000 hours of force-synchronised recordings across multiple manufacturing setups.

The consequence is that the supervision signal knows about tolerance. A policy graded on task completion cannot distinguish a clean insertion from one that jammed and got forced. A policy graded on whether the measured wrench matched the predicted one can, because those feel different.

Results

Metric	Facet-0	Strongest baseline
Mean success, five sub-millimetre computer-assembly tasks	82%	15%
Placement precision	0.5 mm
Command latency	50 ms
Training data	ManuFacet-1K, 1,000 hours force-synchronised

A five-fold gap usually means the baseline was doing something structurally wrong rather than the new method being tuned better. If your only signal is task completion on a task where most of the information lives in the contact, you are learning from the wrong channel.

What came before

Episode three covered SoftVTBench, which found that on deformable objects a real fraction of rollouts scored as successes had crushed the thing they held. A success label cannot see a tolerance violation, because the label is a bit and the violation is a continuous quantity nobody recorded. Facet-0 is the constructive answer. Predict the force, and tolerance stops being an unobserved variable.

The 50 ms latency deserves a mention on its own. Force feedback only helps if you close the loop faster than the contact event evolves, and hitting that with a learned policy is often the difference between a method that ships and one that stays in a paper.

Limitations

This is one group, one dataset, five tasks, no independent reproduction, and the strongest baseline is not named in the abstract. Every number here is the authors' own, on the authors' own setups, which is the normal situation for a week-old preprint and still worth saying before quoting a five-fold gap. Whether force prediction generalises past tasks defined by contact is unknown. My guess is it degrades toward "no worse than a normal policy" on free-space tasks, where the expected wrench is near zero and the critic has nothing to discriminate on. That is a guess, and somebody should run it.

Further reading: - Facet-0, the action-wrench pairing, the critic and the ManuFacet-1K dataset - One Demonstration, Many Objects, generalising manipulation through local contact geometry, from the same week - VeriPhy, the adjacent argument that a scalar score cannot tell you which physical obligation was violated or when

Waymo and Tesla ran opposite experiments

Waymo opened three cities on 1 September and ran two of them exclusively on the minivan it built to be cheap, of which roughly 300 exist. Two days later Tesla started charging passengers for rides in the Cybercab, a purpose-built two-seater with no steering wheel, of which 45 are authorised in Texas. Both called it scaling. They are answering different questions, and the number that settles it is published by neither.

This story has nothing to do with the ownership thread running through the rest of the week, and it is strong on its own numbers.

Disclosure: I lead AI at Kodiak, which operates in autonomous trucking. PlusAI, mentioned below, is a competitor.

Waymo's Ojai

Item	Figure
New cities, 1 September	Denver, San Diego, Tampa
Total US cities	14
Total fleet	4,000+ vehicles
Ojai fleet	~300
Cities where the Ojai is the only option	Denver and San Diego
Paid rides per week	500,000+ (separately reported), targeting 1 million by year end

The Ojai is the detail worth your attention, a minivan carrying Waymo's sixth-generation driver and designed to be cheaper to build, operate and maintain.

The conservative play with a cheaper platform and a large installed fleet of expensive Jaguars is to migrate the cities you already run, where you have mapping, depots and a known operational profile, and where a new vehicle's problems surface against a familiar baseline. Waymo did the opposite and opened brand new markets exclusively on the cheap car. That is a stronger signal about unit economics than any number the company published.

Tesla's Cybercab

The Cybercab is a two-seater with no steering wheel and no pedals, running AI4 compute and camera-only FSD. On 3 September, Tesla started charging the public for rides in Austin, which makes it the first vehicle designed from scratch without driver controls to carry paying passengers in the US. That is a real milestone. Removing the controls is a regulatory event as much as an engineering one, and somebody had to go first.

The number underneath is 45, the count of Cybercabs authorised for commercial robotaxi use in Texas on launch day, against a Waymo fleet above 4,000. Tesla also charges for driverless Model Y rides in Austin, Dallas, Houston, Miami, Orlando and Tampa, so the Cybercab is an addition to an operating service rather than the service itself. The launch event was closed, which surprised a community expecting an unveiling.

The number neither publishes

Waymo is proving its vehicle is cheap enough by opening cities on it. Tesla is proving its vehicle exists at all by putting paying riders in it. Scoring them on fleet size flatters one, and scoring them on hardware novelty flatters the other.

In a year the metric that decides this is cost per mile, fully loaded, including depot, cleaning, remote assistance, insurance and vehicle amortisation. Neither publishes it. Waymo has at least built a vehicle around the number, and Tesla's camera-only architecture is a bet on the same number from the sensor side. Until one of them has a reason to disclose, everything else is a proxy.

Further reading: - TechCrunch on the three Waymo launches, city list, fleet size and the Ojai exclusivity - Motor1 on the Cybercab launch, the vehicle, the 45 authorisations and the paid-ride milestone - Forbes on the closed launch event, and the community reaction to it

Also this week

Language models still cannot do geometry. Alibaba's Qwen team shipped Qwen-Drive-1.0, its first driving foundation model, unifying 3D perception, driving VQA and motion planning. The architectural tell is that it keeps an external bird's-eye-view head for detection and map segmentation, plus a separate Planning Expert for trajectories. A major open-weights lab entering driving resets what counts as an academic baseline, and it took 370 upvotes on Hugging Face, the most of any embodied paper this window.

Three answers to the data question, in one week. ZimaBlue says scrape and pretrain, scaling to more than 120,000 hours of human and robot egocentric video and moving real-robot zero-shot success from 36.1% to 77.8%, with a Slow-Fast architecture that runs 30 Hz action prediction on a single RTX 4090. RoboTok says index and retrieve, learning a latent motion space over 3D hand trajectories in actor-centred frames so the same motion is findable whether the camera is on your head or across the room. Kinetic Blocks says buy it. All three are pipelines of things going right, and episode two's FACT result showed that a world model trained only on successes will show you success under an action that should break the task. The data market got liquid this week without acquiring a category for failure.

A world-model audit that ships an interface. VeriPhy compiles a prompt into typed physical obligations before any frame is observed, then gates calls to frozen experts, with each verdict resolving to supported, contradicted or unknown with full provenance. On 149 clips carrying 304 annotated flaw records it detects 228 failures against 222 for just prompting a monolithic backbone, a gap of only six. The authors publish that anyway and argue the score was never the point, which I respect and which is a hard sell to a field that rewards leaderboards.

Money. Anthropic's IPO prospectus is expected after 7 September, on Q2 revenue of $10.9B and a first operating profit around $559M. Unitree fell to 546.51 yuan, more than 50% below its debut high, on fundamentals that were all in the prospectus. PlusAI is going public via SPAC at an $800M pre-money on $25M of revenue. And the largest cheque in the window went to surgery, $700M from Medtronic into Cornerstone Robotics, which quietly runs the deployment economics the humanoid sector keeps promising.

This week in one table

Item	What it is	Link
NVIDIA acquires Hugging Face	$12.93B, $11.9B to shareholders plus ~$1B retention, close expected H1 2027, no mention of robotics anywhere in the post	NVIDIA
Deal confirmation and terms	Structure, Delangue on approaching Huang, other unnamed bidders	TechCrunch
Practitioner reaction	Nithya Ruff (Linux Foundation), Duane O'Brien (OSI) and David DeSanto (Anaconda) on record; discovery neutrality named as the live risk	VentureBeat
Isaac GR00T 1.7 on LeRobot	The July 2026 partnership that makes the acquisition a robotics story, with Cosmos 3 announced as coming	NVIDIA
Kinetic Blocks marketplace	Graded, licensed, chain-of-custody humanoid data in LeRobot v3.0 format, proprietary KBQS score, gated beta from 1 September	humanoid.guide
Skild AI S1	Video in the context window replaces fine-tuning; 66% vs 9% on unseen tasks, ~380x demo multiplier, no paper	Skild
S1 coverage	Independent write-up of the S1 announcement	The Robot Report
Temporal robustness	Expert and ACT both 100% at nominal speed; 84% vs 53% at maximum speed; 35 of 47 failures are insertion misalignments	arXiv
Facet-0	Actions paired with expected wrench, scored by an Action-Wrench Critic; 82% vs 15% on five sub-millimetre assembly tasks	arXiv
Peg-in-Bench	Fully 3D-printable modular insertion benchmark, configurable tolerances, STL files on GitHub	arXiv
ZimaBlue	120,000+ hours of embodied video, 36.1% to 77.8% zero-shot, 30 Hz on an RTX 4090	arXiv
RoboTok	Internet-scale retrieval for human demonstrations via actor-centred 3D hand trajectories	arXiv
VeriPhy	Typed physical obligations compiled before any frame is seen; 228 detections against 222 for monolithic prompting	arXiv
Knowing When to Stop	Training-free action-chunk truncation from cross-attention entropy, on π₀.₅ and X-VLA	arXiv
Qwen-Drive-1.0	Alibaba's first driving foundation model, with a separate BEV head and a separate Planning Expert; 370 HF upvotes	arXiv
LightNav-0	4B generalist navigation model, first on 4 of 8 benchmarks, 67.4 complete-set average, beating an 8B specialist	arXiv
Matrix-Game 3.5	Parameter-free geometry-aware patch memory for streaming interactive world models, minute-long horizons at 720p	arXiv
H3-World	A video generator turned into a controllable world model with 0.199% trainable parameters and 8,000 gameplay samples	arXiv
WorldReward	VLM reward model judging action consistency and visual quality jointly, used for RL post-training	arXiv
Puffin-World	Unified perception, generation and reconstruction with native 3D world state; 0.29 degree median roll error	arXiv
Humanoid Safe Stop	A learned value function for whether the robot can still stop safely, the whole-body analogue of braking distance	arXiv
One Demonstration, Many Objects	Generalising manipulation from a single demo through local contact geometry	arXiv
Lucida	Real-to-sim scene modelling split into parse, generate and place, producing composable rather than monolithic scenes	arXiv
HitMem	Hierarchical temporal 3D memory with context-aware retrieval for dynamic environments	arXiv
DSG	Dynamic 3D scene graph construction for embodied agents in changing indoor environments	arXiv
Loco-manipulation on humanoids	A system for fast, resilient and adaptable loco-manipulation behaviours	arXiv
Non-Prehensile Throwing	Throwing without grasping, framed as a reinforcement learning problem	arXiv
Waymo opens three cities	Denver, San Diego, Tampa; 14 cities, 4,000+ vehicles, ~300 Ojai minivans exclusive in two of the three	TechCrunch
Tesla Cybercab	First purpose-built no-controls robotaxi carrying paying riders in the US; 45 authorised in Texas	Motor1
Cybercab launch event	Closed rather than public, to the surprise of the community	Forbes
Zoox	Begins supervised testing in Houston and San Diego with retrofitted Highlanders	TechCrunch
PlusAI SPAC	Autonomous trucking public at $800M pre-money on $25M revenue, factory-built trucks targeted for 2027	The Robot Report
Lyte Series C	$165M at $1.6B for coherent 4D vision silicon that measures motion instead of inferring it	The Robot Report
Robot.com and Sodexo	Seven-year deal, 500+ sidewalk robots, 2.4M tasks; the CEO says advertising revenue is part of the model	The Robot Report
Medtronic and Cornerstone	$700M plus ex-US distribution rights for the Sentire surgical system	The Robot Report
Gravis Robotics (17 Aug, catch-up)	$200M Series A led by SoftBank for a hardware-agnostic excavator retrofit; largest Series A in construction robotics	Bloomberg
Anthropic IPO prospectus	Expected after 7 September; Q2 revenue $10.9B, first operating profit ~$559M	The Motley Fool
Unitree drawdown	546.51 yuan, more than 50% below the 1,100-yuan debut high	Bloomberg
Locus Robotics Series G	$41.6M from Tiger Global, Goldman Sachs AM, G2 and Scale Venture	FinSMEs
Reframe Systems	$40M to scale robotic microfactories for home building	The Robot Report
Enovis and eCential	$180M acquisition in surgical robotics	The Robot Report
FPGAs as a security enforcement point	Hardware root-of-trust below the model, the silicon-side echo of Anthropic's Model Hardware Standard. Episode listing only, no numbers published	The Robot Report
VLAct (27 Aug, out of window)	Representation-centric continued pretraining beats a full-data baseline on 20% of trajectories and 16 GPUs	arXiv

What we're watching

Who measures download-ranking parity on Hugging Face, and how often? The observable test of Huang's compute pledge is whether artifacts not tuned for NVIDIA hardware rank and download like the ones that are. A baseline collected before the 2027 close is worth far more than one after.

Is the robot ecosystem incidental to a $12.93B platform bet, or unmentioned on purpose? NVIDIA's announcement never says robotics, LeRobot, Isaac or physical AI while the trade press leads with all four. The antitrust review runs into 2027, which gives both readings a motive.

Does a proprietary grade improve robot data, or only price it? Kinetic Blocks scores every dataset with KBQS. The tell will be whether buyers publish their own acceptance criteria alongside the score, or whether the number becomes the standard.

Does anyone outside Skild reproduce in-context task specification at long horizons? An independent group, on named hardware, with a named baseline, would settle in a week what the blog post cannot.

How many published success rates were measured at exactly one execution speed? Two policies scoring identical 100% diverged by more than thirty points when hurried. The follow-up is whether speed is the only hidden axis or the first one anybody bothered to vary.

Is force prediction a general fix, or a contact-task fix? Facet-0 goes from 15% to 82% where contact defines the task. Whether it transfers where the expected force is near zero is the experiment I most want somebody outside the group to run.

When does a robotaxi operator have to publish cost per mile? Waymo opened two cities exclusively on a vehicle built to be cheap. Tesla launched 45 purpose-built ones. Neither discloses the number, and the likely forcing function is a public market rather than a regulator.

Which axis is compute-efficient at the frontier, scale or representation? ZimaBlue gets 36.1% to 77.8% from 120,000 hours of video. VLAct beats a full-data baseline on 20% of the trajectories and 16 GPUs. The answer may flip with embodiment.

And one I am retiring. After four episodes, nothing has moved on whether physical prompt-injection defenses like text masking survive warehouse tasks that require reading real labels. It comes off the list unless somebody commissions the experiment. If you have run it, tell me.