The empty square: forty-eight profiles and what each one actually shares

On 30 October 2025, twenty-four authors posted a paper in which a Unitree G1 humanoid and an Agilex dual-arm robot worked the same order: "I'm hungry and order a normal burger." The twenty-four are at Peking University, the Beijing Academy of Artificial Intelligence, the Chinese Academy of Sciences' Institute of Automation and Beihang University. The two machines split the order by planning against one shared store the authors call STEM, a Spatio-Temporal-Embodiment Memory that "integrates spatial scene geometry, temporal event history, and embodiment profiles into a shared representation". Two different bodies, one memory, one job.

That is the closest public thing to what this site says nobody sells. Two of our pages carry a sentence a reader could try to falsify. Crew technology says that of the forty-eight profiles we surveyed, "none of them sells a site-level memory that a mixed robot fleet reads and writes in near real time". Where value sits says it in six words — nobody sells a shared robot brain. The posts carrying the working came down on 11 September 2026. Here it is again.

The finding is narrower than an absence. Robot fleets share exactly four things in systems that exist: model weights, offline; intent and traffic state; operations data; and language at task boundaries. The fifth square — one site memory that a mixed team reads and writes while it works — is empty, for two commercial reasons rather than a technical one. A hardware vendor has no reason to build the layer that coordinates its competitors' machines, and the companies with the models do not operate sites.

What the survey was

On 5 September 2026 we went through the field company by company: twenty profiles in the United States, twenty-eight in China, forty-eight in all. Behind them: roughly 120 web searches and 85 full-page reads, resting on 187 public source URLs — 86 in the US pass, 101 in the China pass. Two notes for a reader who counts. Those forty-eight profiles name fifty-three companies, since four of them bundle two or three firms — Diligent with Serve, Dyna with Genesis AI and RLWRLD, Dorabot with Youibot, Ant Group with Robbyant [inference]. The split is by research pass rather than by domicile. Dexmate is a Santa Clara company inside the China file; Sharpa is Singapore-headquartered; RLWRLD is Korean, sits in the US file, and is the one name with no public source behind it.

The US pass: Figure AI, Agility Robotics, Boston Dynamics, Apptronik, 1X, Tesla Optimus, Physical Intelligence, Skild AI, Google DeepMind, NVIDIA, Diligent Robotics with Serve Robotics, Dexterity AI, Field AI, Amazon, Gecko Robotics, Sanctuary AI, Generalist AI, Dyna Robotics with Genesis AI and RLWRLD, Locus Robotics, Symbotic.

The China pass: Unitree, AgiBot, UBTech, Galbot, Fourier, Xpeng, Xiaomi, LimX Dynamics, Booster Robotics, Astribot, Leju, Dexmate, EngineAI, Kepler, Standard Robots, Siasun, Geek+, Hai Robotics, Pudu, Keenon, Dorabot with Youibot, X Square Robot, Ant Group with Robbyant, Alibaba, ByteDance Seed, BAAI, Huawei Cloud CloudRobo, Sharpa.

Family one: weights, offline

Shared: a policy checkpoint. Timescale: minutes to nightly. Hardware assumption: one embodiment per checkpoint.

The measured exemplar is a paper, not a product. Learning While Deploying is arXiv 2605.00416, from the Shanghai Innovation Institute, AGIBOT Finch and Columbia University, 1 May 2026. It runs sixteen AgiBot G1 dual-arm robots collecting rollouts while a central learner trains on them and "broadcasts the updated shared policy to the robot fleet every 50 training steps". On-robot control runs at 30 Hz. The fleet gathered roughly 60 robot-hours of online data inside a four-hour wall-clock budget, across eight real tasks of three to five minutes each, at 95 percent average success.

The line that matters is the system description, not a result: "edge actors upload complete episodes, a centralized learner fetches versioned replay data." Episodes go up, weights come down. The robots share no map, no object state and no observation of each other. The best-measured fleet-learning result we found is an upload-and-redistribute loop [inference].

The sold versions publish no mechanism: Figure describes a dock-time upload and a redistributed policy [from 5 Sep baseline, not re-verified this pass]; Agility publishes "regular software updates". And one correction we owe our own notes: Figure's two-robot case is the original Helix of 20 February 2025, coordinated "through natural language prompts", and neither Helix post claims zero communication between two robots. That gloss was ours.

Family two: intent and traffic state

Shared: where each robot intends to go, plus lift and door state. Timescale: path-planning seconds. Hardware assumption: a fleet adapter per vendor.

Open-RMF, which arrives in Singapore public healthcare as RoMi-H, is the open and deployed exemplar, and it draws its own boundary: "All fleet managers that are integrated into RoMi-H must report the expected itineraries of their vehicles to the traffic schedule." Itineraries, lifts, doors. No perception, no object maps, no event history. RMF is a broker rather than a brain, "an intermediary to communicate and negotiate between standalone systems", each fleet still under its own proprietary manager. Locus LocusONE and Geek+ RMS ship the same shape and share work state, not memory [from 5 Sep baseline].

The largest learned version is Amazon's DeepFleet, and it is not for sale. The paper, arXiv 2508.08574, describes four architectures: Robot-Centric at 97M parameters trained on about 5 million robot-hours, Robot-Floor at 840M, Image-Floor at 900M and Graph-Floor at 13M. Its metrics are Dynamic Time Warping distance and Congestion Delay Error. The number everyone quotes needs care, our own pages included. Amazon's blog says DeepFleet "will coordinate the movement of robots across our fulfillment network, improving the travel time of our robotic fleet by 10%", on the page announcing its millionth robot across more than 300 facilities. That sentence is forward-looking, and the paper reports no travel-time figure at all [single source: Amazon blog; a projection, not a published result]. Our Crew page already states it that way, and this is the working behind it.

Family three: operations data

Shared: telemetry, missions, maps and site history, through a console a human reads. Timescale: the dashboard refresh. Hardware assumption: the vendor's own fleet.

The only family whose exemplars are unambiguously products. Boston Dynamics Orbit is "Shared orchestration and intelligence software for all your Boston Dynamics robots". It gives "a unified view of robot activity, site performance, and fleet health for Spot, Stretch, and eventually Atlas", and the on-premise option is real: a 1U Site Hub. Agility's Arc "connects Digit to your existing warehouse automation—including AMRs and management and execution systems". Neither page claims shared memory or shared learning, and neither says what happens when the console is unreachable.

Formant, InOrbit and Viam sell the same shape independent of any robot maker. The one figure in circulation for Formant, about US$250 per robot per month, comes from a competitor's comparison post [secondary: competitor comparison blog]; Formant's own pricing page returns a 404. In the one family that is a product category, nobody publishes what it costs.

Family four: language at task boundaries

Shared: a sentence. Timescale: the handoff. Hardware assumption: both robots can reach the same cloud model.

Gemini Robotics ER 2, published 30 July 2026, reports 91.3 percent accuracy and a 0.96 second mean absolute distance on moment finding [single source: vendor]. Its multi-robot demonstration has robots using "shared semantic understanding to handoff and complete complex tasks", on Apptronik's Apollo 2 and a Franka F3 Duo. Two facts matter more than the benchmarks. It is cloud only and in preview, and its model card bars use in "healthcare, transportation, or other areas where safety protocols are vital, and a malfunction could reasonably foreseeably lead to death, personal injury, or property damage." The first hospital that wants this cannot have it, by the publisher's own terms.

UBTech's BrainNet 2.0 is the only group-brain claim from a humanoid manufacturer, and it rests on two press accounts with a single origin. Sina Finance, 24 July 2025, says in Chinese that BrainNet 2.0 and Co-Agent "form the AI dual-loop system of the industrial humanoid robot", with robots as dynamic nodes. A paid press release of 3 March 2025 names BrainNet as "a software framework for humanoid robot collaboration" and does not mention 2.0 at all [single source: two press accounts, one origin]. No architecture paper, no network latency figure, no statement of what is shared between the Walkers, no third-party verification. A claim of exactly the right shape with no evidence attached. Our own product page cited the second of those for an illustration the release does not contain; we read the URL this week, found it absent, and took the claim off the page in both languages.

Ops data

• Boston Dynamics OrbitBoston Dynamics robots only; on-premise Site Hub
• Agility Arcbuilt around Digit; also speaks to third-party AMRs
• Formant, InOrbit, Viamvendor-neutral; none publishes a price

Intent and traffic

• Open-RMF / RoMi-Hitineraries, lifts, doors — no perception
• Amazon DeepFleetAmazon-internal, not sold

Weights

• Figure, Agility: fleet software updatesshipped; no mechanism published [5 Sep baseline]
• AgiBot Learning While Deployingpaper; 16 robots, new policy every 50 steps

Language

• Gemini Robotics ER 2cloud only, in preview; model card bars healthcare
• UBTech BrainNet 2.0two press accounts, one origin; no architecture published

Memory

• RoboOS-NeXT / STEM, BAAIpaper, Oct 2025; not sold, code not released

• A site can buy it or run it today
• It cannot: internal to one operator, in private preview, or a paper

The shape across the four families matters more than the families themselves [inference]: the further you move from telemetry toward cognition, the less of it is a thing you can buy. Operations data is a product category with price lists. Traffic is an open standard plus one internal system at Amazon. Weights are sold with the mechanism undisclosed. Language is in private preview and excludes healthcare. Memory is a paper.

The fifth square, stated so it can be checked

Here is the definition we checked against, at the level our Brain platform page publishes. A site memory holds four things: a fused scene graph with object permanence, so a thing seen by one robot stays known when no robot is looking at it; an episodic log with provenance; the site's semantic knowledge, meaning zones, exception policies and standard operating procedures; and a versioned skills library.

It has to guarantee three things over those four. Consistency rules for many writers, because a memory with several authors fails in four specific named ways rather than degrading gracefully — the subject of a companion article. A local replica on each robot, so the memory is not a single point of failure between a machine and its next move. And a published degraded-mode contract: what a robot does when the shared memory is unreachable, how fast it notices, and what it stops doing.

1. What it holds

   Fused scene graph with object permanencea thing one robot saw stays known when none is looking at it

   Episodic log with provenancewhat happened, and which machine says so

   Semantic site knowledgezones, exception policies, standard operating procedures

   Versioned skills librarythe skill, and which version of it ran

2. What it must guarantee over them

   Consistency rules for many writersa memory with several authors fails in four named ways

   A local replica on each robotso the memory is not a single point of failure before the next move

   A published degraded-mode contractwhat a robot does when the memory stops answering, how fast it notices, what it stops doing

   Nothing in this survey publishes the third. RoboOS-NeXT does not say whether STEM runs in the cloud or on site; Gemini ER 2, Orbit, Arc and BrainNet 2.0 publish no offline behaviour.

The three hold over all four: the arrow reads 'over', not 'then'.

The honest exception

Back to the burger. RoboOS-NeXT is the one thing in the corpus that could falsify us. A planner sits over STEM, low-level controllers under it. The authors evaluate 200 tasks in each of three domains, 600 in all. Adding robots helps: Average Execution Steps per Task falls from 34.8 with one robot to 14.7 with three and 8.5 with five, while success slips from 76.6 to 69.7 percent. Mixed teams work too: two humanoids with two quadrupeds reach 10.5 AEST at 70.7 percent success, a humanoid with two wheeled robots 16.2 at 72.5 percent. In the household scenario without injected errors it succeeds 89.2 percent of the time. AEST is a step count, not a duration; a reader who assumes seconds will misread all of it.

All of them come from simulation. Section IV-A says so: "we conduct experiments in a mock setting that abstracts away physical uncertainties and focuses on system effectiveness." The real-robot work is a separate section, three demonstrations on a Unitree G1, an Agilex dual-arm and a Realman arm, with no trial counts, no success rates and no run durations. The paper is titled "lifelong". What ran on hardware was three scripted scenes.

So the word in our sentence is "sells", and it holds in three parts.

Nobody sells it. Four institutions wrote the paper, and the code would land in BAAI's repository. BAAI is a non-profit research institute whose project site and FlagOpen repository carry no offering, no pricing and no support. It is BAAI's to release, and BAAI has not released it. You cannot download it either: the public repository is Apache-2.0 and contains RoboOS 1.0 and 2.0 only, with 2.0's release still pending on its roadmap. Neither STEM nor NeXT is in it, and the project page does not mention NeXT, as of 11 September 2026.

What is in that repository needs saying, because a reader following our link reaches it before we mention it. The RoboOS project page describes version 1.0 as a cloud-based embodied brain with distributed cerebellum modules and "real-time shared memory", and 1.0's code is there, Apache-2.0, today. That is the nearest thing to a falsifier in existence, and we would rather name it than have it found. It does not reach. No vendor stands behind it — no price, no support, nobody to call. Version 2.0, described in the repository as a "lightweight single-machine deployment" on a Redis master-slave store, is still listed as unreleased. The two limbs below hold against it as they hold against everything else here: nothing published shows either version running a mixed team at site scale for weeks, and neither publishes a degraded-mode contract. A framework you can clone is a real thing, and worth more than a press release. It is not a product a hospital can buy, and it is not STEM.

None of it has run on a mixed team at site scale for weeks. No experiment in the paper lasts hours, days or weeks; the sequence-length variable counts tasks, not time; no run duration is disclosed.

And nobody publishes a degraded-mode contract. RoboOS-NeXT does not say whether STEM runs in the cloud or on site, or what a robot does when the memory is unreachable; we looked for that specifically. Gemini ER 2 publishes no offline behaviour. Orbit ships an on-premise Site Hub and still does not say what a Spot does when the hub is down. Arc publishes none. UBTech publishes none. That limb is scoped to the systems in this survey.

Somebody has built a shared robot memory and demonstrated it on three real robot platforms. You cannot buy it, the code you can download is a different and smaller system, it has never run on a mixed team at site scale for weeks, and neither it nor any product here says what a robot does when the memory goes away.

Why the square is empty

The first reason is not our inference. It is published by the consortium that had to build around it: Open-RMF's FAQ, explaining why each vendor keeps its own fleet-management infrastructure, says vendors "may view it as an unacceptable liability to share computing resources with another vendor." That is the argument, stated as an engineering constraint by the people who wrote the workaround.

The product pages say it again in their first lines: Orbit is for "all your Boston Dynamics robots", BrainNet is a Walker-fleet story. Each is a reason to buy more of that vendor's hardware, not to coordinate a competitor's machine in the same corridor.

Agility's Arc is the partial exception, and we should say so: we quoted it above, and our own Crew page says it speaks to MiR and Zebra fleets. Its full sentence is "connects Digit to your existing warehouse automation—including AMRs and management and execution systems" — AMRs being, precisely, other vendors' robots. But read the direction of travel. Arc connects other fleets to Digit; it does not offer to hold their memory. A vendor will talk to a competitor's machine when that makes its own robot more useful. Hosting the store that machine thinks with is a different offer, and it is the one nobody makes.

The second reason is that the companies with the models do not operate sites. Gemini ER 2 cannot run on premise, and bars healthcare. Physical Intelligence and Skild are partner-gated and not self-hostable [from 5 Sep baseline]; Skild's S1 page offers early access with no pricing and no claim about experience travelling between robots. NVIDIA's fleet story is training, simulation and workload orchestration, with no shipped runtime fleet-memory product [from 5 Sep baseline, not re-verified this pass] — the weakest limb here, and the first place to look to falsify this. BAAI publishes papers, and has not published this code.

We re-checked the perishable part on the day of writing, against three indexes: The Robot Report's feed, DeepMind's blog and Physical Intelligence's. Between 5 and 11 September 2026 none of them carried a shared-memory or fleet-brain product announcement. DeepMind has published nothing on robotics since July 2026, Physical Intelligence nothing since π0.7 on 16 April 2026, and Skild's S1 was announced in August 2026, not September — the September item was NVIDIA writing about it. Zero announcements in a week, and that will stop being true without warning.

What we are doing, and what we are not claiming

Tidewell Crew is not shipped and the site memory service is in design. The empty square is a survey result, not a description of something we have filled, and the latency figures on our pages are design targets until a test date sits beside them.

We are not the only people who could build this layer. What we claim is that we will write down what it has to guarantee before we sell it — the four contents, the consistency rules, the per-robot replica, and the degraded-mode contract nothing in this survey publishes.

That last one is the question to put to a vendor. Not whether their robots share a brain, which everyone will now say yes to. Ask what a robot does in the ninety seconds after the shared memory stops answering, ask for it in writing, and ask what happens to the machines on site that are not theirs. Forty-eight profiles in, we read five of them closely enough to answer that question on their behalf, and none of the five has the document. The bar is one paragraph high — ours is already on the Crew page, and its numbers stay design targets until a test date sits beside them. The next article in this pair is about what breaks when somebody writes it.