Atlas Loses Its Pinky, Only One Humanoid in Eight Walks

Boston Dynamics' hand team spent a working day with their pinkies taped to their ring fingers, came out of it convinced the finger wasn't worth its size, cost and power, and on 1 October shipped Atlas a hand with four digits, thirteen degrees of freedom and pressure sensing across every fingertip and the palm. Every reason the company gave was about building the thing, repairing it or simulating it, and none was about something the hand could do that a five-fingered one couldn't. The same week an investment bank put a number on how far that trade has already spread. Of roughly 31,000 humanoids produced this year by the leading makers, BTIG counts 13% that walk on two legs. The industry is quietly subtracting human anatomy, and it is changing what it counts at the same time.

Atlas's new hand

The new Atlas hand has thirteen degrees of freedom against seven in the previous generation, spread over four digits. Actuation is direct-drive, with one encapsulated actuator type sitting inside the joints, no tendons and no exposed cables, and each actuator module can be replaced on its own. Dense pressure sensing covers the fingertips and the palm. Boston Dynamics says the hand is slightly larger and stronger than a human one, shows it carrying a loaded minifridge over 100 lb, and says it is designed to be built at 100,000 units a year. The published reasoning is manufacturability, repairability and simulation fidelity. Capability is not on the list.

Degree of freedom (DOF): one independent way a mechanism can move. A hinge that only bends is one DOF. Counting DOF is the crudest measure of a hand's dexterity, and the field quotes it because it is the easiest thing to count.

What is in the hand

Four digits and thirteen DOF is more dexterity than the hand it replaces, so the headline reads backwards. Boston Dynamics moved almost twice as much articulation into one fewer finger, and spent the volume the pinky freed up on actuators it can seal, model and swap.

Why the tendons went

Most dexterous robot hands are tendon-driven. Motors sit in the palm or the forearm and pull cables routed through the fingers, which is how you get strong, slim digits without stuffing a motor into every knuckle. The cost shows up elsewhere. Cables stretch and abrade, friction through a routed path depends on the pose and on how worn the sheath is, and the relationship between motor command and fingertip force drifts with use. For a lab hand that is an annoyance. For a hand you intend to make in six figures and service in the field, it is a maintenance contract.

Direct drive puts the actuator at the joint and deletes the cable. That costs volume at the joint, which is where the pinky's three degrees of freedom went. What Boston Dynamics buys is a joint whose torque is a function of current rather than a function of current and accumulated wear, and a module a technician can pull without restringing anything.

Designing for the simulator

Boston Dynamics says the hand was designed for high dynamic fidelity simulation, with domain randomisation and actuator proprioception, so it can train grasping with reinforcement learning in simulation and move the policy to hardware. Human demonstrations go in on top of that.

Sim-to-real: train a policy in a physics simulator, then run it on the robot. It works when the simulator's errors are smaller than the policy's tolerance for them, and it fails in ways that look like the robot being confidently wrong.

Designing hardware so the simulator can model it is a different posture from designing hardware and then fighting the simulator. In my own work on perception for highway autonomy, the parts that survived contact with sim-trained policies were the ones whose behaviour was a clean function of a measurable state. Tendon friction never is. Boston Dynamics appears to have let the training method pick the mechanism, which says more about where it thinks the hard part lies than the press material does.

The pinky

A day of work with the pinkies taped down convinced the team that three more degrees of freedom were not worth the size, cost and power. Alberto Rodriguez, Boston Dynamics' director of robot behavior, told IEEE Spectrum: "Hands are a ruthless design trade-off. There's no way around it, you're always giving up on something."

I like the experiment and I want to be clear about what it is. One team, one day, no task battery, no recorded data, no published results. It is a vivid design heuristic and a long way short of evidence. Its conclusion agrees with every other constraint in the build, which is both why it carries weight and why I am suspicious of it.

What is missing

No success rates. No tactile benchmark. No comparison against another hand on any shared task. Everything above is a specification and a set of demonstrations, and a demonstration of a hand rolling golf balls tells you the controller worked when the video was shot. The 100,000 units a year is a design target, not a forecast. Boston Dynamics builds Atlas in low volume today.

The interesting unanswered question is the data one. The field's two-year argument for human form factor is that a human-shaped hand lets you retarget human demonstration video onto the robot. Boston Dynamics says it trains from human demonstrations and also removed a finger, and neither the Spectrum piece nor The Robot Report's write-up says how the retargeting handles the mismatch. There are reasonable answers. Nobody published one.

And the pressure sensing is the largest tactile collection apparatus anyone has proposed. The largest public tactile dataset today runs to roughly 30,000 hours, the field says it wants 100,000, and what exists is spread across about 21 mutually incompatible sensor formats. Whether any of Atlas's touch data gets published, and in which schema, matters more to the rest of us than the finger count.

Further reading - IEEE Spectrum on the hand, with the Rodriguez interview and the design reasoning. - The Robot Report's version, which is tighter on the actuation and serviceability details.

What the humanoid production count counts

BTIG broke down roughly 31,000 humanoids produced this year by the leading makers, Unitree and AgiBot among them, and found 32% are full-size and 13% are bipedal. That is around four thousand units with legs, out of a figure the sector quotes as evidence of industrial adoption. The headline count was never wrong. It was answering a narrower question than the people repeating it assumed.

Every valuation in this sector runs through a shipment figure that three organisations could not agree on within 50%, and the open question has been who audits it. BTIG leaves the buyer question alone and answers the adjacent one, which is what shape the units are.

The subtraction is worth doing. If every bipedal unit is also full-size, which is the charitable reading, then roughly six thousand full-size humanoids produced this year are standing on wheels or a base. The volume underneath all of it is half-size machines going to education and research, carried by Chinese models in the $4,900 to $21,500 band. BTIG also notes that Tesla's Optimus has roughly 10,000 unique parts, a serviceable explanation for why full-size ramps are slow.

Two honest caveats. The underlying BTIG note is not public, and both the GuruFocus and Seeking Alpha write-ups quote the same three figures, so treat 31,000 / 32% / 13% as single-sourced until the methodology appears. And it circulated through investment press, which is why the robotics commentariat has not picked at it yet. Hold the numbers loosely and the direction firmly.

The cheap end agrees

Three machines launched inside about a week make the same argument from the price side.

Flourish makes the argument cleanly. Users teach it a chore by guiding it through the motion in a phone app, no code, in about thirty minutes, after which the company says it repeats and schedules the task with no remote operator and stops the arms when a person comes within reach. Demonstrated chores are picking up shoes, clothes and toys, wiping tables, fetching, watering plants, loading dishwashers and clearing litter boxes. Flourish Robots is split between San Francisco and Paris, led by Antoine Marcel, with pre-seed money from Families Fund, and the unit is reported to be Raspberry Pi-based.

$3,555 sits an order of magnitude below 1X's NEO at $20,000 or $499 a month. The specification is the whole argument. Three and a third pounds of payload, wheels, two basic grippers. My read is that the price came from deleting capability rather than from a breakthrough, and for picking toys off a floor that is a sensible trade. It becomes a bad trade the moment anyone quotes Flourish 1 in the same breath as a machine meant to work a shift.

Further reading - BTIG via GuruFocus, the breakdown and the Optimus parts-count aside. - Flourish 1 at Interesting Engineering, with the full specification. - CSRC's three criteria for humanoid IPOs, where a securities regulator decides the brain or the hands is what counts as proprietary.

Simple-WAM, and what one denoising step buys

Tsinghua's Leap Lab rebuilt the two competing designs for world-action models on a matched backbone, matched data and a matched compute budget, which nobody had done. On the tasks they were trained on, the fast design and the slow one tie. Off those tasks the fast one collapses, scoring 5.9% against 69.9% at learning a new task from action-free video and 10.0% against 68.9% on a held-out task on a real robot. The whole gap traces to the first denoising step. Their method keeps that step and throws away the rest of the picture, reaching 73.6% and 87.8% on those two axes at 74.7ms per action chunk against 286.9ms, with code and weights released.

What a world-action model is, and why anyone bothers

Robot policy: takes a camera image, the robot's joint state and a text instruction, and emits the next short burst of motion, usually called an action chunk.

World-action model (WAM): a policy that also learns to predict what the camera will see after that motion executes.

The reason to predict video is supply. There is vastly more video of the physical world than there is teleoperated robot data, and if a model learns how objects fall, roll, slide and deform by watching, you need far fewer hours on a real arm. I have believed for a while that this is the right architecture to bet on. A model that learns how the world behaves and then learns to act in it is a different object from one that memorises a map from pixels to joint angles.

The frames come out of a video diffusion model, which starts from pure noise and cleans it up over a series of steps until an image appears. That process is denoising, and it is expensive, because each step is a full forward pass through a large network. So the field split.

Explicit WAM: denoises actual future frames every time the robot acts. Slow, because inference pays for the full denoising chain.

Latent WAM: learns to predict the future during training, then skips the prediction at run time. Fast, and the claim was that nothing important is lost.

Fast-WAM made the latent case in March, and it was persuasive. Train with the imagined future, act without it, keep the accuracy and lose the latency. Anyone who has shipped a model onto a vehicle or a robot wants that bargain, because inference budget is the constraint that binds hardest in deployment.

What the matched comparison found

Leap Lab, with USTC and Beijing Institute of Technology, built both paradigms on a Wan 2.2 backbone interpolated to a 1024-dimension ActionDiT, with matched data and budget, and evaluated three axes of generalisation on LIBERO, LIBERO-Plus and RoboTwin. Their baselines are named FastWAM and FastWAM-Joint, so this is a deliberate reckoning.

The third row is the one that ends the argument. You show the model ordinary video of a task it was never trained on, with no robot actions attached, and then ask it to perform that task. The explicit model does it roughly seven times in ten. The latent model does it roughly once in seventeen. Fast-WAM was right about everything it measured, and the shortcut fails the moment the robot meets something new, which is the only moment that matters outside a benchmark.

The mechanism

Leap Lab traced nearly the entire gap to the first denoising step. The explicit model gets its benefit from beginning to work out the future and almost nothing from finishing the picture. Their phrase for it is preparing the future rather than generating it.

Simple-WAM takes that literally. It gives the model a set of fully noised video tokens as placeholders for the future, runs a single forward pass over them alongside the action prediction, and never renders a frame. It also adapts the training noise schedule to match, because a model trained to denoise over many steps is not calibrated for a single one. That is the whole method, and it beats both predecessors on all three axes at 3.8x the explicit paradigm's speed.

The ablation is what makes the claim stick. Replace the noised video tokens with learned query tokens, or with zeros, and every generalisation score collapses. The noise itself is doing the work.

Here is my reading of why, and I could be wrong about it. To take even one step from noise toward a future frame, the model has to commit to what is about to happen in the scene. That commitment is the computation the action head needs. Everything after it is detail about pixels, and the robot never needed the pixels. If that holds, the field spent two years paying for rendering in order to buy a representation that was already there at the first step.

Limitations

Most of these numbers come from simulation benchmarks. The real-robot evidence is one task on one platform, and 10.0% against 87.8% on a single held-out task is a vivid number with a thin denominator behind it. The paper is days old and nobody has rebutted it, which is some distance from nobody being able to. It drew 123 upvotes on Hugging Face daily papers, among the highest for a robotics item in the window, so attention is not the problem.

I take it seriously anyway because of the shape of the evidence. Two rival designs, one backbone, one budget, three axes, and an ablation that tries to break the authors' own explanation. I would rather an architectural argument ended this way than by whichever camp ships the larger model.

Further reading - What Makes World Action Models Generalize?, the paper, with the full tables. - Simple-WAM code and weights from LeapLabTHU. - UniWAM, which reports a log-linear scaling law for unified human-robot co-training and represents low-level actions in natural language, running against the instruction-binding result below.

Nucleus's 40%, and DYNA's intervention metric

Nucleus published close to two hours of uncut footage of its humanoids working in a German factory, hesitations and human interventions included, and its founder volunteered that the demonstration was about 60% autonomous and 40% teleoperated. The company did not say how it computed the split. Days earlier DYNA Robotics argued the field should quote Mean Time Between Interventions instead of per-task success rates, and published no MTBI of its own. Add the per-task success rates Figure and Sunday Robotics argued over last week, and the field has three ways of counting and still no table.

What Nucleus showed

Founder and CEO Melvin Schwarz posted the footage on 1 October: parts picking, shelf loading, product handling and cart transport, with the pauses and the human takeovers left in rather than cut out. He told Humanoids Daily the demonstration is roughly 60% autonomous and 40% teleoperated, and that the robot regularly runs four-to-six-hour shifts completing a job start to finish, which is not a claim of four to six hours without intervention.

The hardware is modified Unitree G1 platforms with custom end effectors and protective suits, inside a major German industrial customer. Nucleus bills robotics-as-a-service by the hour for completed work and says it exited stealth with a working factory deployment on day 91. Its founders come from CERN, ESA, 1X, NEURA and Agile Robots.

For a long time I have been asking who publishes the number that makes them look worse. The answer turned out to be a startup running on somebody else's robot, and the number is 40%.

Two caveats, and both outlets that covered it flagged them unprompted. The split describes this demonstration and not the company's robots in general. And Nucleus has not said how it derived the percentages, so nothing in the video lets anyone check them. Call it transparency. Verification would need somebody else's stopwatch.

Nucleus also argues the teleoperation is the strategy rather than an embarrassment. It treats task prompts, expert instructions and teleoperation corrections as training signal, recording interventions alongside observations. That is a coherent position, and it is also the position that makes a 40% number safe to publish, so I would not read the disclosure as pure candour.

DYNA's Taku and Mean Time Between Interventions

Mean Time Between Interventions (MTBI): how long a robot runs before a person has to touch it. A reliability metric rather than a capability metric.

DYNA Robotics launched Taku, a wheeled semi-humanoid with two 7-DOF arms taking either parallel-jaw grippers or dexterous hands, a folding lower body for reaching high and low, and four steerable wheels. It runs DYNA 2.1, which the company describes as its proprietary DYNA 2 world-action model under a vision-language orchestrator and above a whole-body controller, trained on what DYNA calls a million hours of human and robot data. The demonstration is an hour-long commercial laundry shift, running washers and dryers, folding towels, sorting by size, shelving at varying heights and recovering from its own errors. The founders are Lindon Gao, York Yang and Jason Ma, with CRV and First Round behind them.

The million hours is unaudited, and roughly eight times the largest comparable figure I have on record, ZimaBlue's 120,000 hours. Deployment counts in hotels, laundromats and restaurants are undisclosed.

The metric argument is the part worth keeping. Commercial laundry is the same work Sunday Robotics used recently to attack Figure's home-chore numbers, after Figure reported 40% on toy tidying under whole-chore scoring and Sunday answered with 99.1% on laundry scored per garment. Those two percentages cannot sit in one table, because the scoring unit differs. DYNA stops arguing about the denominator and changes the measure, and MTBI is the first metric proposed this year that a facilities manager would buy against. Shift length divided by interventions is a number you can write into a service contract.

It is also a number DYNA has not published. So the field now has a company that published an autonomy split without a method, a company that proposed a metric without a value, and before them a Figure and Sunday argument over success rates that measure different events. Three ways of counting, and the table still does not exist.

Further reading - Nucleus's footage, covered by Interesting Engineering. - DYNA Robotics on Taku and DYNA 2.1. - Input perturbations change how VLAs succeed, the research-side version of the same complaint, showing two models with matching success rates behaving very differently on the trajectories they succeed on.

California SB 951, the Worker Technological Displacement Act

Governor Newsom signed thirteen more AI bills on 30 September, and the coverage went almost entirely to SB 947, the robo-bosses bill. The one that changes what anybody can verify is SB 951. A mass-layoff notice caused in whole or substantial part by an AI system or other automated technology must now be headed "This notice is for a technology displacement", describe the job functions being automated, and name the category of technology responsible. The Employment Development Department publishes the notices and a quarterly statewide summary. The phrase doing the work is "or other automated technology", because it pulls a physical robot in alongside software.

What Cal-WARN is

Cal-WARN: California's version of the federal Worker Adjustment and Retraining Notification Act. Larger employers must give advance written notice of a mass layoff, relocation or termination. It is a disclosure regime, not a ban. Nothing in it stops a layoff; it makes the layoff legible and on a clock.

SB 951 (Gómez Reyes) bolts a technology disclosure onto that existing machinery, which is why it will work. No new agency, no new filing. The notice already had to go in. Now it has to say what replaced the people.

One date caveat matters if you go reading about this bill. Early coverage describes 90 days' notice and a 25-worker threshold, which was the introduced version. The bill was narrowed before enactment, and the enacted text carries the Cal-WARN thresholds above.

Why this is a robotics story

Every humanoid pitch for three years has rested on a labour number nobody outside the company could check. The claims go into press releases and come back out in valuations, with no public counterfactual anywhere in the loop.

From 2027 California manufactures one. A warehouse with a few hundred staff that lets sixty pickers go because it installed robotic arms clears both thresholds, has to file, has to head the notice as a technology displacement, has to describe the picking function as automated, and has to name the technology category. The state then puts it on a list and sums the list every quarter.

It will be a messy dataset. Employers write their own technology category, "in whole or substantial part" is going to be litigated, and any firm that can plausibly blame demand instead will. Layoffs below the thresholds are invisible, and so is the commoner pattern where nobody is fired and a vacancy is never filled. Attrition does not file a WARN notice.

Even allowing all of that, this is the first dataset in existence that lets anyone check a vendor's labour claim against a filing. Nobody covering the signing framed it that way, because the story ran as workplace-software regulation. The companion bills point the same direction. SB 947 (McNerney) bars relying solely on an automated system to discipline or terminate someone, and AB 1331 and AB 1883 restrict workplace surveillance, including banning it in bathrooms.

My own view is that this is close to the right shape for AI regulation. It leaves the deployment alone and asks an employer to write down something it already knows, which costs an honest one almost nothing. The usual failure mode here is a rule that stifles the research and leaves the harm untouched. This one manages the opposite.

Further reading - SB 951, enacted text. - The Governor's office on the thirteen-bill signing.

Instruction-action binding

A Taiwanese group argues that robot policies ignore the instruction because the imitation objective never forces them to use it, and reports a fix that takes a real UR5e from 8% to 88% on unseen positions. That lands eight days after RoboFollow concluded the opposite, having watched four separate fixes fail. Two groups a week apart agree on the diagnosis and disagree about the prognosis.

Hung-Jen Chen and colleagues start from a familiar embarrassment. Fine-tuned pi-0.5 and GR00T-N1.7 clear 90% in-distribution and then fail under counterfactual changes that demand a different action. Their behavioural analysis is the useful part. Failed rollouts do not flail. They retain the source behaviour, or cleanly switch to another task the model was shown during training. The policy is retrieving a trajectory, not reading the sentence.

They name that instruction-action binding and argue the objective permits it by construction. If, in the whole training set, no two demonstrations ever share an instruction while requiring different actions, then a grounded policy and an instruction-keyed lookup table are indistinguishable to the loss. Nothing in training ever prices the difference.

Equivariant Counterfactual Training supplies the missing pressure. It adds demonstrations where the same instruction requires different actions in distinguishable scenes, and pairs each demonstration with its counterpart inside the same gradient update, so the model cannot satisfy one without separating them. The authors also report that CALVIN five-task completion improved with the loss alone.

The 8% to 88% is one task, one arm, one lab, and the authors' own. 36% to 59% on the simulated position swap is the sober number, and it leaves the problem a long way from solved. The mechanism claim is what makes this worth tracking. If the failure is a property of the training distribution rather than the language interface, anyone willing to collect paired counterfactuals can fix it cheaply. The test that settles it is somebody else's arm in somebody else's lab, running RoboFollow's own protocol against a policy trained this way.

Further reading - When Instructions Retrieve Trajectories. - RoboFollow, the counterpart that says four fixes fail.

Anthropic's prospectus, and the price of memory

Reporters have now read Anthropic's draft IPO prospectus, all 261 pages of it, including roughly 80 pages of risk factors against 48 describing the business. Across every outlet that covered it, nobody reports a mention of robotics, embodied AI, actuators or the Model Hardware Standard the company previewed five weeks earlier. Meanwhile Micron said on 30 September that the memory shortage worsens through 2028 with over 75% of fiscal 2027 output already committed. In this week's financial story, the physical layer of AI appears only as a cost line.

The reported numbers come from a draft that can still change, and the S-1 is not on EDGAR as of 3 October, so nobody outside the draft's readership can search it.

Disclosed through the same document on 1 October, Broadcom will provide up to $42B in convertible notes against Anthropic's TPU lease obligations, covering about a third of a $125.2B five-year commitment. The chip designer finances the lease of its own silicon, in paper that converts into the lessee's equity. The $125.2B is the number to stare at. One software company has committed more to compute leases than the entire industrial robotics industry turns over in a year.

The absence of robots is the part I care about, and an absence is weak evidence. Reporters covering an IPO look for revenue, losses and risk, and a two-paragraph product line would never survive the edit. Treat it as reported-but-unverified. When the public document lands, the word is either in there or it isn't.

Micron's shortage is the concrete half. CEO Sanjay Mehrotra said the company cannot say when supply catches demand, expects conditions to worsen over two years even with cleanroom space arriving from late 2028, and has committed more than 75% of planned fiscal 2027 output. Customers will pay much higher prices next year. The mechanism is HBM capacity for AI accelerators crowding out conventional DRAM.

Nobody will file that as a robotics story and it is one. Every on-robot policy worth running is memory-bound rather than FLOP-bound. FLUX 3 Action, the 7B open-weights policy released last month, needs 32GB in BF16 and 24GB at FP8 to run at all, and NVIDIA's Jetson T4000 ships 64GB. A two-year DRAM squeeze lifts the bill of materials on every edge-inference robot, in the same seven days three companies announced robots whose entire pitch is the price.

Further reading - TechCrunch on the draft prospectus. - Broadcom's $42B against the TPU lease. - Micron on DRAM through 2028.

This week in one table

What we're watching

• Who publishes an intervention rate with the method written down? Nucleus put out a 60/40 autonomy split without saying how it was computed, and DYNA proposed Mean Time Between Interventions without publishing one. The first company to publish a rate alongside its definition of an intervention will tell us more than any demo reel. The harder question is whether a customer ever makes one a contract term, because that is the point at which the number stops being marketing.
• Do the large world-action models adopt a single-forward-pass future? Simple-WAM says the useful object is the noised future rather than the rendered one, and the code and weights are public. Watch for two things: whether anyone reproduces the latent-WAM collapse on a second real robot, and whether the physical-fidelity audits move in the same direction as generalisation when they are re-run on a Simple-WAM-style policy, or in the opposite one.
• Does anyone split the humanoid count by buyer? BTIG split roughly 31,000 units by form and found 13% bipedal, from a note that is not public. Nobody has separated a robot sold into a factory from one sold to a subsidised training centre. If an analyst or the IFR prints that breakdown, the question is whether any humanoid valuation gets re-marked when they do.
• What is in California's first quarterly displacement summary? From 2027 the EDD publishes technology displacements with the technology category named by the employer. How many notices name a physical robot rather than a software system is the first real measurement of what these machines have replaced, and whether any other state copies the requirement determines whether it stays a California curiosity.
• Does any of Atlas's touch data get published? Boston Dynamics is building a hand for 100,000 units a year with dense pressure sensing on every fingertip and the palm, which is the largest tactile collection apparatus anyone has proposed. Whether it releases any of it, and in which of the field's roughly 21 incompatible sensor formats, decides whether this becomes a public resource or a private moat.
• Does Equivariant Counterfactual Training hold on a second embodiment? RoboFollow watched four fixes fail and concluded the problem might be the interface. A Taiwanese group reports 8% to 88% on one real UR5e. One of those is going to look naive in six months, and the experiment that decides it is another lab, another arm, and RoboFollow's own protocol run against an ECT-trained policy.
• When the S-1 lands, is robotics in it? Anthropic's draft has been read in detail by reporters and the public filing is still not on EDGAR. Neither robotics nor the Model Hardware Standard appears in any coverage. Once the document is searchable, the absence becomes checkable, and whether a driver spec for operating a microscope or a robot arm survives inside a company on a quarterly clock is the follow-up.
• Which sub-$30,000 robot's price moves first? Micron has committed more than 75% of fiscal 2027 output and says the shortage worsens through 2028. Three companies launched robots this week whose entire proposition is the price. One of those bills of materials moves before the others, and the useful disclosure would be on-robot inference cost per task, before and after.