Policies
One foundation model, a few long-horizon chores, tested in rooms the robot has not mapped. Data is the argument. 56% is not a household product.
Science desk · Robotics · Status 2025–2026
The parkour reel is no longer the whole story. In 2025–2026 a handful of companies moved humanoids from lab clips into early factory and logistics pilots, published learned whole-body policies that survive a new room, and started talking like manufacturers. Batteries, hands, safety paperwork, and the long tail of messy scenes are still the hard parts. Pair this with the parent robotics stand on the science hub and the live science news desk. Industrial arms still earn the rent. This page is about the two-legged bet.
Three public facts, not a slogan. First: learned controllers — vision–language–action stacks, diffusion-style action models, whole-body policies — left papers and entered company reports you can read. Second: a few platforms are on real floors for tote work, sequencing, or inspection, with named customers and hour or tote counts the companies themselves publish. Third: factories for the robots, not just for the cars or totes they might one day move, are under construction or already assembling units. That is a different year than the one where every humanoid story was a backflip.
It is not a solved product category. General-purpose household autonomy is still a research and teleop story. Multi-shift deployments that survive months without a hero video remain the watch item on the science hub. Company press, supplier gossip, and drone footage of steel frames can all outrun a verified shipment. This essay keeps those layers separate on purpose.
One foundation model, a few long-horizon chores, tested in rooms the robot has not mapped. Data is the argument. 56% is not a household product.
Electric industrial humanoids with published lift, battery-swap, and IP ratings — plus a warehouse biped sold as cage-optional. Specs are brochures until a floor keeps them.
First assembly halls and converted auto lines. Capacity numbers in the millions are plans and chain reports, not a delivery table.
Figure’s Helix 2.5 note (September 17, 2026) is the cleanest public card on learned whole-body generalization this season. The company says it pretrained a foundation model on Index, its human-behavior dataset, adapted that single base to three behaviors — tidying a living room, folding towels, making a bed — and then ran the robot in 30 Bay Area homes with no data collected in any of them and no fine-tuning on those rooms or the objects being handled. “Zero-shot” here means the evaluation houses and objects, not the tasks. The tasks were specified with fine-tuning data gathered elsewhere.
The comparison they published is the part worth keeping. Holding task-specific data, architecture, training, and evaluation fixed, a policy trained from scratch succeeded on 9% of zero-shot trials. The Index-pretrained policy succeeded on 56%. Success, they say, required the whole task: every toy in the basket, every towel folded and placed, or the whole bed made — no partial credit. That is a company evaluation, not a third-party audit. It is still a useful sentence: broad human video, in their report, is doing most of the generalization work.
Helix 02, on their telling, could already coordinate a whole body over long horizons — dishwasher unloading, a logistics task they say ran autonomously for 200 hours — but those systems learned from data collected where the robots would operate. Helix 2.5 is the “new house, same policy” claim. They also say the pretrained model matched a Helix 02 in-environment success rate with about half the task-specific data, and that doubling Index data improved downstream robot-action prediction smoothly enough to call it a scaling law. Those are their plots. Treat them as a published experiment, not as proof that homes are a finished market.
The rest of the Helix note is a data-scale pitch: Index generating “roughly 35 minutes of new human experience every second,” and a committed $3.5 billion of compute. Those are company figures. They belong next to the parent desk’s reminder that robot data is still scarce next to internet text, and that sim-to-real gaps do not vanish because a scaling curve looks smooth on a held-out loss.
Boston Dynamics’ Atlas product page is no longer a research-mascot site. The electric machine is framed as an industrial humanoid for material handling: barcode workflows, fleet learning through their Orbit software, and a battery the robot is said to swap itself so a shift can continue. Published specs, as of the page checked in September 2026:
The same page says Atlas is already in a customer facility with Hyundai for field testing on real-world sequencing tasks, and that a later “product version” incorporates lessons from that testing. The commercial path they sketch is early adopters first: evaluate, train on the job, then adopt. That is a pilot-and-pipeline story, not a claim that Atlas fleets are a normal warehouse SKU. Treat the kilograms and the IP rating as manufacturer specifications. Field testing is the interesting verb.
Agility Robotics unveiled Digit 5 on September 15, 2026. The pitch is specific: a humanoid engineered for “cooperatively safe work at scale,” meant to work near people without the physical cages traditional industrial robots use. The company says this sits on more than 65,000 hours of Digit 4 operation, including named sites at GXO, Schaeffler, Amazon, and Toyota Motor Manufacturing Canada, and a 100,000-tote milestone at GXO’s Flowery Branch facility near Atlanta with “approximately 98% accuracy while on-task.” Those hours and that tote count are Agility’s. They are the longest industrial tenure story a humanoid maker is currently willing to put on a press page.
Hardware upgrades they list: a new leg and cycloidal actuators for repeated lifts up to 50 lb (22.7 kg); a 90-minute battery they say charges in 9 minutes (a 10:1 run-to-charge ratio, up from Digit 4’s 2:1); swappable grippers on ISO-standard flanges; 5 ft 11 in (1.81 m) tall, 284 lb (129 kg), reach to 7.2 ft (2.2 m). Safety architecture, in their telling: people-detection that can avoid, stop, or sit; visual and auditory motion cues; an independent safety controller. They say Digit was the first humanoid to pass an independent field evaluation on a customer line for industrial safety standards administered by OSHA — cite that as Agility’s claim, not as a government endorsement you can look up in a single OSHA bulletin from this page.
Scale language needs a pencil. Agility says Digit is assembled at RoboFab, a 70,000-square-foot plant in Salem, Oregon, “designed to produce up to 10,000 Digit robots per year” at full capacity. As of May 2026 they report more than $300 million in multi-year Digit 5 orders “subject to satisfaction of certain contractual milestones.” Early access is expected in the first half of 2027, with general availability by the end of 2027, and a first commercial path outside North America into the EU and UK, CE mark expected. The same release carries a long forward-looking-statements block. Designed capacity and milestone-gated orders are not the same as 10,000 robots on docks.
Tesla’s own words are narrower than the supply-chain week. The Q2 2026 shareholder update (July) says the company decommissioned Model S and Model X lines at Fremont, is installing first-generation Optimus lines, and expects to start production soon — “in anticipation of production in 2026.” Initial builds “will be used in our Optimus Academy for training data collection and further functionality development.” The manufacturing table lists Optimus in California and Texas as Construction, with capacity marked “-”. Texas is “building construction now in full swing.” That is the primary card: a converted Fremont line, an internal academy, a Texas building, no installed annual capacity in the table, no customer shipment count.
September 2026 secondary reporting runs ahead of that table. CnEVPost (September 21), relaying The Paper and 21jingji, says a Tesla robotics team is auditing Yangtze River Delta suppliers — Tuopu, Sanhua Intelligent Controls, Joyson Electronics — for a third-generation Optimus designed for volume, and that those outlets describe Fremont as a dedicated line with a planned annual capacity of 1 million units and a Texas line with a long-term designed capacity of 10 million, “expected to start production around summer next year.” Those million-unit figures are industry-chain numbers in a secondary story, not a Tesla delivery report. Drone watchers have been filming steel at Giga Texas; that is construction progress, not a first-year output forecast. Treat Gen 3 mass-production talk as prep: audits, line conversion, a building going up. Do not cash it as proven shipments.
Unitree is the research-volume door you can actually click. The G1 product page lists a starting price of US $13.5K (tax and shipping excluded), about 2 hours of battery life, and 23–43 degrees of freedom depending on the EDU configuration. The same page tells individual buyers the global humanoid industry is “in the early stages of exploration” and to keep a safe distance. That is a catalog robot for labs and developers, not a named factory fleet in Unitree’s own spec sheet.
Apptronik Apollo is the enterprise-plus-contract-manufacturer story. Apptronik and Jabil’s February 25, 2025 release announced a pilot and strategic collaboration: Jabil as worldwide manufacturing partner, and Apollo units to be integrated into specific Jabil operations — including, they write, the lines that will build Apollo. That is a real named partnership. It is still a collaboration announcement, not a 2026 unit-shipment table.
1X Neo is the home-preorder narrative. 1X’s October 28, 2025 launch and order page offer Early Access at $20,000 or a $499/month subscription, with U.S. deliveries said to start in 2026. The same pages are explicit that complex chores Neo does not know can be handled by a scheduled 1X “Expert” who remotely supervises — teleoperation as a product feature, not a rumor. Engadget’s contemporary write-up matches that disclosure. A preorder with a remote human in the loop is not household autonomy. It is an honest, and limited, commercial shape.
The parent robotics stand already had this right. Hardware still bites. A 90-minute Digit pack or a 4-hour Atlas pack is a shift-planning problem, not a footnote. Hot-swaps and 9-minute charges are how you paper over energy density; they do not invent a new battery chemistry. Hands and tactile sensing still lag cameras. Figure’s chores and Digit’s tote history are impressive inside their published rubrics. A messy, novel object in a cluttered aisle is the everyday remainder.
Safety and liability are catching up in public. Agility says it is contributing to ANSI/A3 TR R15.108 (a technical report for dynamically stable industrial mobile robots, including humanoids, in the U.S. and Canada) and to ISO 25785-1, which they call the first international safety standard for the humanoid category. Those are standards in motion, not a finished certificate you can hang on every two-legged machine. Learning controllers make the paperwork stranger: who is responsible when a policy does something nobody wrote as a branch? This page will not answer that. It will say shipping into human spaces is a legal and insurance problem, not only a demo problem.
Reliability is the difference between a pilot and a product you can forget about. Helix 2.5’s 56% is a whole-task, no-partial-credit number on three behaviors. Digit’s 65,000 hours are the other pole: narrower work, longer clock. Neither is “the robot did the house, then the warehouse, then went home.” Watch multi-shift deployments that survive a quarter without a new video. Watch whether fleet learning closes the data gap without papering over a safety case. The live desk will catch the headlines. This page is here so the headlines have somewhere honest to sit.
You do not need a tour badge to notice the geography. Figure’s Helix 2.5 evaluation used 30 unseen Bay Area homes. The company’s manufacturing and HQ story in 2025–2026 sits in San Jose — local real-estate reporting put a new headquarters plant at 3960 North First Street after a Sunnyvale chapter. Tesla is converting Fremont’s former Model S / Model X space into Optimus lines. Agility opened a Fremont, California facility in July 2026 as a hardware and physical-AI hub, while Digit assembly stays at RoboFab in Salem. The South Bay is not “where humanoids were solved.” It is where several of the loudest English-language programs collect data, convert floors, or keep an office. The same valley still runs on ordinary industrial arms. That split is the local sentence.
Keep going
Parent stand
Foundation models meet factory floors — capabilities, limits, and what to watch next.
Ticker
Headlines and the breaking strip. This essay is context, not a feed.
Status
Nanomaterials, robotics, space — public-knowledge altitude for 2025–2026.
Sibling essay
Another nested science desk: cited, local, allergic to press kits.
Disclaimer. Educational science literacy, not investment advice, not a product review, and not safety, legal, or insurance guidance. Figures and quotes come from the public pages listed above, checked in September 2026. Company evaluations, manufacturer specifications, order backlogs, designed plant capacities, and supply-chain reports can all outrun independently verified deployments and shipments. Operations, specs, and timelines change — trust the outbound pages over a static paragraph. For workplace safety or procurement, use current standards, counsel, and qualified engineers.