US robotics players, September 2026

Three marks run through this post and mean what they meant in the research file. [single source] is a claim found in exactly one place. [secondary] is an aggregator or a blog rather than a primary source. [inference] is our own reasoning rather than a sourced fact. Where a company discloses nothing, the table says "not disclosed" rather than guessing.

The pass ran on 5 September 2026: about 45 web searches and 40 full-page reads across company pages, SEC filings, earnings transcripts, The Robot Report, TechCrunch, Forbes, CNBC, the FCC and Commerce, Gartner and Forrester, and a nurses' union.

The company tables use one set of row labels: (a) product, price and business model; (b) brain architecture; (c) on-robot compute; (d) fleet and multi-robot, including shared memory; (e) memory and continual learning; (f) verified deployments; (g) documented weaknesses; (h) what the company explicitly will not do or does not sell.

Figure AI (Figure 03, Helix 02)

Field	Finding
(a) Product / price / model	Figure 03 (launched 9 Oct 2025): die-cast, soft textile shell, palm cameras, fingertip tactile (3 g threshold), 2 kW inductive charging via feet, 10 Gbps mmWave data offload; BotQ initial capacity 12,000/yr, target 100,000 over four years (Figure 03 page). Commercial model: RaaS reported at about US$1,000/robot/month [single source, secondary: TSG Invest/eesel]; no official price. Valuation about US$39B [secondary].
(b) Brain	Helix 02 (Jan 2026): S0 = 10M-param whole-body controller at 1 kHz trained in 200k+ parallel sim envs, replaces 109,504 lines of C++; S1 = 200 Hz visuomotor transformer over head and palm cameras, tactile, proprioception; S2 = VLM at 7 to 9 Hz for language and scene. S2 writes a latent to shared memory that S1 consumes (Helix page). Longest demo: 4-minute, 61-action autonomous task. Nothing published on cloud dependence; the design implies fully on-robot inference.
(c) Compute	Original Helix: "dual low-power embedded GPUs", model-parallel S1/S2 (Helix page). Figure is on NVIDIA's Jetson Thor early-adopter list (NVIDIA PR, Aug 2025) but Figure has not published the Figure 03 SoC.
(d) Fleet / shared memory	Claim: robots "upload terabytes of data for continuous learning" while docked (Figure 03 page). Adcock's fleet-learning description is local experience, then central cloud validation and aggregation, then a redistributed policy; an independent review notes this takes hours to days, not real time (ETC Journal, Jun 2026). No shared site memory product.
(e) Memory / continual learning	No persistent episodic or site memory disclosed. Learning is offline retraining. No learn-from-correction feature described.
(f) Deployments	BMW Spartanburg (Figure 02, single unit): 1,250+ h over about 6 months, 10-h shifts Monday to Friday, 90,000+ parts, 30,000+ X3s, 84 s cycle, over 99% placement per shift; the forearm was the top failure point (Figure "production-at-bmw" page; BMW press). BMW moved Figure 03 to Spartanburg logistics sequencing in 2026 (BMW press, Jun 2026). BMW chose Hexagon AEON, not Figure, for its first German deployment (Leipzig; BMW press). Second customer "reportedly UPS" [single source, secondary]. Logistics demo: 4.05 s/package, about 20% faster within months [secondary]. Home: 8 cleaning skills demoed Mar 2026 [secondary].
(g) Weaknesses	Whistleblower suit (Robert Gruendel, principal safety engineer) filed N.D. Cal. Nov 2025: robots "powerful enough to fracture a human skull", a quarter-inch gash in a steel fridge door during a malfunction; Figure denies, case ongoing (CNBC). No uptime or MTBF disclosed. The single-unit BMW pilot is the only audited-scale number. Lost the German BMW slot to Hexagon.
(h) Won't do	Fully vertically integrated; dropped OpenAI for in-house Helix; no evidence of licensing Helix to third-party bodies or of any partner-body programme [inference from strategy; no explicit quote found]. No consumer sales channel yet.

Agility Robotics (Digit, Arc)

Field	Finding
(a) Product / price / model	Digit (v5 next): 16 kg (35 lb) tote capacity, 4-h battery, human-centric form (Agility solutions page). RaaS, described as an "industry-first" RaaS humanoid deal with GXO. Going public via Churchill Capital XI SPAC at US$2.5B pre-money, over US$620M gross proceeds (SEC Form 425, Jun 2026).
(b) Brain	Not published as a named model. Collaboration with Google DeepMind mentioned in the SPAC deck; Agility is a Gemini Robotics 2 "supporting partner" (DeepMind blog). NVIDIA Halos launch partner (full-stack safety system).
(c) Compute	Jetson Thor early adopter (NVIDIA PR).
(d) Fleet / shared memory	Agility Arc is a cloud automation platform: facility mapping, workflow definition, fleet orchestration, connects Digit to AMRs, WMS and WES, KPIs including uptime, throughput and MTBI (SEC 425; solutions page). Orchestration and monitoring only; no claim of robots sharing learned experience or memory.
(e) Memory / continual learning	Not disclosed. "Regular software updates" over the air only.
(f) Deployments	9 customer facilities, more than 65,000 operating hours; customers GXO (Spanx facility, Georgia, paid since Jun 2024, 100,000+ totes), Schaeffler, Toyota Motor Manufacturing Canada (3-robot pilot to 7 Digits commercial, Feb 2026), Mercado Libre; more than US$300M multi-year Digit v5 orders (about 1,000 robots in backlog per management [single source]); 30+ company pipeline; RoboFab up to 10,000/yr, about 75% US-sourced parts (SEC 425). OSHA-recognised NRTL field inspection Nov 2025 against ANSI/RIA R15.08 and ISO 13849 [secondary].
(g) Weaknesses	No public uptime, MTBF or incident numbers (Arc tracks them, they are not disclosed). 4-h battery. GXO's chief executive (Aug 2026): zero humanoids in production in 2026, no ROI yet, about 2 years away (GXO Q2 call; Fortune). Claims of "1,200 Digits deployed" circulating on ifactoryapp contradict Agility's own 9-facility figure; treat as false [secondary, unreliable]. Revenue and net loss are not in the summarised filings (the DRS/A is too large to parse).
(h) Won't do	Policy paper: opposes "weaponization" of humanoids; favours tariffs and anti-dumping measures on Chinese robots and "voluntary standards" over regulation. No third-party brain licensing; the software is Digit-only [inference].

Boston Dynamics (Atlas, Stretch, Orbit, DeepMind)

Field	Finding
(a) Product / price / model	The electric Atlas product version was unveiled at CES on 5 Jan 2026; all 2026 units are committed to Hyundai RMAC and Google DeepMind (Automate.org). Price: "below cost of two US manufacturing workers for two years", about US$320k ceiling per sources (KED Global, Jan 2026) [single source]; analyst guesses US$130–145k [secondary]; no public purchase until 2027. Stretch: purchase or RaaS not disclosed; Orbit plus Scout software about US$15–25k/robot/yr [single source, secondary: Komo]. Installed base 2,000+ robots [secondary].
(b) Brain	Atlas: Large Behavior Models with Toyota Research Institute (language-conditioned end-to-end policies) plus Google DeepMind Gemini Robotics models for perception and reasoning (Robot Report; the DeepMind blog lists Boston Dynamics as a partner). Orbit uses Gemini-powered VLMs for inspection analytics (Orbit page).
(c) Compute	Atlas: Jetson AGX Thor T5000 (NVIDIA PR; Robot Report).
(d) Fleet / shared memory	Orbit is "shared orchestration and intelligence software" for Spot, Stretch and "eventually Atlas": mission scheduling, multi-site dashboards, Site View 360-degree history, anomaly alerts, APIs and webhooks, CMMS and WMS integration; deployable on AWS cloud, on-prem as a 1U "Site Hub", or as a VM; SOC 2 Type II (Orbit page). It shares maps, missions and data, not learned skills or episodic memory.
(e) Memory / continual learning	Orbit stores site history (Site View). No robot-level continual learning product.
(f) Deployments	Stretch: DHL MOU for 1,000+ additional units (May 2025); DHL had 10 Stretch across 3 countries as of Jun 2025 [single source]; up to 700 cases/h; customers DHL, Maersk, Gap, H&M, Otto/Hermes, Arvato, NFI (Stretch page); 20+ facilities [secondary]. Atlas: RMAC and DeepMind only.
(g) Weaknesses	Stretch cannot handle all loads, for example canoes and ironing boards (DHL). Atlas is unavailable to outside buyers before 2027, and the brain is a Google dependency [inference]. No published reliability numbers.
(h) Won't do	Atlas in 2026 is Hyundai and Google only. Orbit is Boston Dynamics robots only, with no third-party bodies.

Apptronik (Apollo, Gemini Robotics)

Field	Finding
(a) Product / price / model	Apollo 2 (30 Jun 2026): modular, bipedal or wheeled base; the wheeled version is "designed to conform with existing safety standards for industrial mobile robots" (GlobeNewswire PR). Apollo 2 is a data collection and training platform; Apollo 3 is the first true commercial product (2027 per secondary). US$520M Series A-X in Feb 2026 at about US$5B; Series A total over US$935M. Robot Park: about 90,000 sq ft Austin data factory.
(b) Brain	Google DeepMind Gemini Robotics (VLA plus ER 2); Apollo 2 data feeds Gemini. The ER 2 multi-robot demo is Apollo 2 plus a Franka FR3 Duo (DeepMind blog). Sharpa Wave 22-DoF hands on Apollo 2.
(c) Compute	Jetson AGX Orin or Orin NX [secondary]; notably absent from NVIDIA's Thor early-adopter list.
(d) Fleet / shared memory	None disclosed. Fleet learning is Google-side offline training.
(e) Memory	None disclosed; a "combination of teleoperation and autonomous execution" generates the data (PR).
(f) Deployments	Pilots: Mercedes-Benz, GXO (pilot), Jabil (validation tasks: inspection, sorting, kitting). No hours or units disclosed.
(g) Weaknesses	No commercial product until Apollo 3; teleoperation-heavy; the brain is outsourced to Google's cloud ER 2 [inference]; GXO sees no production humanoids in 2026.
(h) Won't do	No brain of its own to license; the hardware is for Gemini.

1X (NEO)

Field	Finding
(a) Product / price / model	NEO home humanoid: US$20,000 purchase (US$200 deposit) or US$499/month, both including "Expert Mode" teleoperation (Robot Report). 30 kg, lifts 70 kg, carries 25 kg, tendon drive, Wi-Fi, Bluetooth and 5G.
(b) Brain	Redwood: about 160M-param VLA at about 5 Hz on an embedded GPU, trained on teleoperated and autonomous episodes; plus an LLM for conversation with conversational memory (Robot Report; 1X Redwood page). Autonomy estimated at 60 to 70% in 2026 [secondary]; the remainder is remote human teleoperators.
(c) Compute	Jetson Thor, described as the "only product... built to support NEO's requirements" (1X GTC 2026 page). Training on HGX B200; Isaac Sim and Isaac Lab.
(d) Fleet / shared memory	Every teleoperation session is recorded and fed to Redwood training ("If we don't have your data, we can't make the product better", per the chief executive). Offline central retraining; no shared memory.
(e) Memory	Conversational continuity and adaptation are claimed; no episodic task memory disclosed.
(f) Deployments	No verified consumer delivery as of 16 Jul 2026; "some this year, some later"; 10,000+ preorders, 10,000/yr Hayward capacity [secondary].
(g) Weaknesses	Teleoperators see inside homes, and the privacy safeguards (blurring, no-go zones, US-based operators) are unaudited company claims; slipping delivery; a 5 Hz policy.
(h) Won't do	Consumer and home only; US and Canada first, international in 2027. No business or hospital offering.

Tesla Optimus

Field	Finding
(a) Product / price / model	Not for sale. Q2 2026 call: the Fremont Model S/X line is converting to Optimus, first runs Q3 2026; V3 at Fremont, V4 at Austin with easier-to-make parts; early units go to an "Optimus Academy" for training data (NotATeslaApp summary of the call). Musk: "hardest product to scale manufacturing that we've ever made"; external sales "as early as" H2 2027 or "by end of 2027" (WEF) [secondary]. Shareholder-letter language softened from "volume production" to "production" between Q1 and Q2 2026 [secondary].
(b) Brain	FSD-derived end-to-end nets; Grok voice; "Optimus Academy" self-play and physics sim; no published architecture.
(c) Compute	AI5 taped out 15 Apr 2026, Optimus and clusters first [secondary]; V4 has "upgraded AI chips", AI5 volume "hopefully next year".
(d) Fleet / shared memory	Musk: fleet data sharing "follows principles used for autonomous vehicles" (Feb 2026 podcast via ETC Journal), which is central retraining, not live shared memory.
(e) Memory	None disclosed.
(f) Deployments	Zero external. Internal units are "primarily for learning and data collection"; "a few hundred" built against 5,000 promised Gen 2 [secondary]; no useful factory work as of Jan 2026 (ETC Journal).
(g) Weaknesses	Timeline credibility; no customer-facing product, no waitlist; hand actuator supply.
(h) Won't do	No external sales channel before 2027; no licensing.

Physical Intelligence (π0.7)

Field	Finding
(a) Product / price / model	π0.7 (16 Apr 2026): a "steerable" foundation model with compositional generalization (TechCrunch). Access is by partnership only; π0.7 cannot be self-hosted (π0 is open via openpi). Reported at about US$300/robot/month subscription [single source, secondary: Black Scarab]. Raised about US$1.1B; valuation US$5.6B, reportedly raising at about US$11B (TechCrunch).
(b) Brain	The VLA family: π0, π0-FAST, π0.5, π*0.6 with RECAP reinforcement learning from its own successes and failures, and π0.7. Hi Robot "inner voice" planner.
(c) Compute	Not disclosed; Physical Intelligence is "evaluating" Jetson Thor (NVIDIA PR).
(d) Fleet / shared memory	None. Hardware-agnostic model licensing; partner deployments in manufacturing, logistics and laundry are unnamed on the site.
(e) Memory	MEM research: short-term is about 2 s of video tokens; long-term is natural-language event summaries up to about 15 min ("I opened the fridge door"); the model chooses what to remember. Within-task only, with no persistence across episodes, days or robots (pi.website/research/memory). π*0.6 ran 5:30 to 23:30 continuous laundry [secondary].
(f) Deployments	Contrary (2026): "no commercial deployment". The Physical Intelligence site says partners are "solving real-world problems" in packaging assembly, laundry and package handling, with no named customers.
(g) Weaknesses	Needs step-by-step verbal coaching for novel appliances; "sometimes the failure mode is... us, not being good at prompt engineering" (TechCrunch). No self-hosting, no site memory, no fleet product.
(h) Won't do	No hardware; no direct end-customer deployments; partner-gated.

Skild AI (Skild Brain, S1)

Field	Finding
(a) Product / price / model	The "Omni-bodied" Skild Brain; S1 (Aug 2026) is an in-context learner: one human video prompt produces up to a 10-min task with no fine-tuning (Skild blog). Not purchasable; no public API or pricing. About US$1.7B raised; about US$14B valuation (Contrary). OEM partners ABB, Universal Robots and MiR. Acquired Zebra's robotics unit (ex-Fetch, which Zebra bought for US$291M in 2021) on 15 Apr 2026, including the Symmetry orchestration platform.
(b) Brain	A single set of weights across quadrupeds, humanoids and arms; trained on teleoperation, human video, sim and glove data ("no golden path").
(c) Compute	Not disclosed.
(d) Fleet / shared memory	Stated goal: "one unified brain" coordinating humanoids, dogs, arms and AMRs via Symmetry orchestration (Skild and Zebra blog). No mechanism for shared memory is published, and orchestration is not shared experience.
(e) Memory	S1: no weight updates at test time; nothing on persistence across sessions.
(f) Deployments	"Limited set of industrial partners", names undisclosed; inherits Fetch AMR customers, count undisclosed.
(g) Weaknesses	66% success on unseen long-horizon tasks at 100k h of pre-training; "for every $1 on data, $3 on QC"; needs video demonstrations rather than language; Pathak says it is "not quite" ready for homes (Robot Report).
(h) Won't do	No robot hardware of its own, although it now owns the Fetch AMR line.

Google DeepMind (Gemini Robotics 2, ER 2, On-Device 2)

Field	Finding
(a) Product / price / model	30 Jul 2026: Gemini Robotics 2 (VLA), ER 2 (embodied reasoning), and On-Device 2, which adapts to a new body with under 200 examples in "a few hours". ER 2 is available via the Gemini API, AI Studio and an Enterprise Agent Platform private preview; the VLA and On-Device are trusted-tester only. No pricing.
(b) Brain	ER 2 is a cloud VLM planner (Gemini 3.5 Flash base), 128k-token input and 64k output; Gemini Live API bidirectional streaming; tool use (VLA, navigation APIs, Google Search); temporal progress tracking (57.4% progress classification, 91.3% moment finding, 0.96 s MAE). VLA success rates 45.7 to 76.3% whole-body, 74.2 to 89.6% gripper, 32 to 92% multi-finger.
(c) Compute	ER 2 is cloud only. On-Device 2 is a local VLA on unspecified hardware.
(d) Fleet / shared memory	Multi-robot collaboration: heterogeneous robots "communicate via a shared semantic understanding to hand off" tasks; the demo is Apollo 2 plus a Franka FR3 Duo; Spot fetch runs via the Orbit and Spot APIs. The mechanism is unspecified, with no statement on central orchestrator against message passing against shared world state. Demo only; no production deployment named.
(e) Memory	Context window only, 128k. No persistent site memory.
(f) Deployments	Partners: Apptronik, Boston Dynamics, Agility, Agile Robots, Franka. No commercial customers named.
(g) Weaknesses	The model card excludes "safety-critical applications" including healthcare and transportation; cloud latency for planning; ER 2 is unavailable on-prem.
(h) Won't do	Sells no robots; excludes healthcare and safety-critical use by terms of service.

NVIDIA (GR00T, Jetson Thor, Isaac, reference humanoid, Cosmos)

Field	Finding
(a) Product / price / model	Jetson AGX Thor became generally available on 25 Aug 2025: dev kit US$3,499; T5000 module 2,070 FP4 TFLOPS, 128 GB, 40 to 130 W, 7.5 times Orin compute (NVIDIA PR). Isaac GR00T N1.7 (early access, commercial licence). Reference humanoid (1 Jun 2026, GTC Taipei): a Unitree H2 Plus chassis plus Sharpa Wave 22-DoF hands plus Thor T5000, 75 DoF, 150 lb; sold by Unitree from about Oct 2026 for academia (NVIDIA PR). Cosmos Reason 2, Cosmos Transfer/Predict 2.5. Halos safety stack.
(b) Brain	N1.7: 3B params, a Cosmos-Reason2-2B VLM plus a 32-layer DiT action head ("Action Cascade"); EgoScale 20,854 h of egocentric human video; the "first scaling law for robot dexterity" (HF blog). Embodiment configs: Unitree G1, AgiBot Genie 1, YAM and others.
(c) Compute	Thor is the de facto humanoid SoC: adopters include Agility, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta; 1X, John Deere, OpenAI and Physical Intelligence are evaluating.
(d) Fleet / shared memory	None. Isaac is a single-robot training, sim and perception stack.
(e) Memory	None at product level; Cosmos Reason is a reasoning VLM.
(f) Deployments	Platform vendor; the reference robot goes to Ai2, ETH, Stanford and UCSD.
(g) Weaknesses	The reference humanoid uses a Chinese chassis while the US FCC banned new foreign humanoid and quadruped imports on 29 Jul 2026 (with exemptions for federal and approved use), which is a tension for US adopters [inference]. N1.7 is humanoid-centric, with no hospital or port workflow layer.
(h) Won't do	Does not sell finished robots; "Android of generalist robotics" (TechCrunch).

Diligent Robotics and Serve Robotics (Moxi)

Field	Finding
(a) Product / price / model	Moxi is a hospital mobile manipulator on an annual subscription, about US$200–400k/yr per hospital [secondary], with a 12-week implementation on existing Wi-Fi. Serve Robotics acquired Diligent for US$29M plus up to US$5.3M earn-out, closed Q1 2026 (Robot Report). Moxi 2.0 rollout on 17 Aug 2026: NVIDIA A2000-class compute, perception 10 to 15 times faster, safety and autonomy onboard so tasks complete without Wi-Fi, 30% faster charging for up to 18 h/day against 9 h before.
(b) Brain	Classical autonomy plus "human-guided learning" of workflows; built with Isaac Sim and Cosmos world models (Moxi 2.0).
(c) Compute	NVIDIA "A2000", RTX A2000-class per Investing.com [single source].
(d) Fleet / shared memory	Multiple Moxi per hospital; Serve intends to apply its 2,000+-robot sidewalk autonomy stack to Moxi. No shared memory claim.
(e) Memory	Learns the hospital map and workflows during setup, and adapts "based on staff interactions" (site page), with no detail.
(f) Deployments	About 100 robots, 25+ hospitals, 1.25M+ tasks (Serve PR); earlier 23 health systems and 31 hospital-level partnerships, 1M+ deliveries, 125,000+ autonomous elevator rides, 20 to 26 min per delivery, 575,000+ staff hours (Robot Report 2025). Moxi 2.0 sites: Endeavor Edward, Providence Saint John's, CHLA.
(g) Weaknesses	MultiCare (Tacoma General, Good Samaritan; 14 robots bought in 2023) discontinued Moxi in 2025: nurses said the robots were "annoying and often got in the way", "needed an escort between floors" and "never delivered meaningful time savings", and nurses had no say in the purchase (WSNA, 2026). The tiny exit value of US$29M signals weak unit economics [inference].
(h) Won't do	US only; no international footprint found.

Dexterity AI (Mech)

Field	Finding
(a)	Mech "superhumanoid": dual-arm on a mobile base for trailer loading and unloading; the Foresight world model (vision, depth and touch) reasons over 3D and time for pack placement. Expanded Kawasaki collaboration (Automate, Jun 2026) on RL030N 8-DoF arms to scale production.
(b)/(c)	Foresight world model; compute not disclosed.
(d)/(e)	None disclosed.
(f)	The FedEx Hagerstown hub expanded from pilot to "significantly larger operational scale" (Jul 2026) after "multiple years" of validation; unit counts and throughput are not disclosed.
(g)	Opaque numbers; a single flagship customer.
(h)	Not humanoid, not general-purpose; parcel-hub focused.

Field AI (FieldFoundation Models)

Field	Finding
(a)	Software-only subscription licensing (integration fee plus recurring), hardware-agnostic from quadrupeds to humanoids; about US$405M raised, over US$2B valuation, Hyundai strategic investment Feb 2026 (Contrary).
(b)	Field Foundation Models: Dynamics FM, Multiagent FM (fleet coordination), risk and uncertainty-aware traversability; EDGE runtime, under 100 ms, fully offline.
(c)	Edge only; hardware unspecified.
(d)	Multiagent FM coordinates fleets; no shared memory claim.
(e)	The models "build context and experience to anticipate outcomes" (marketing), with no persistence details.
(f)	"Several hundred mission deployments across three continents" (Mar 2026); DPR Construction; a Certis partnership (23 to 24 Feb 2026) with a Singapore office, integrating with Certis' Mozart orchestration for security patrols across APAC and Qatar; named partner in Singapore's Punggol physical-AI testbed (JTC, 20 May 2026, with Certis, DHL, Grab, QuikBot, Thoughtworks, Slamtec and Unitree).
(g)	Contrary risks: OEMs may internalise autonomy (Boston Dynamics to Gemini); tariff-driven cost increases made some 2025 deployments "no longer economically viable"; no revenue disclosed. Not manipulation-focused.
(h)	No hardware; no ports or healthcare focus found.

Amazon (DeepFleet, Vulcan)

Field	Finding
(a)	Internal only; 1M robots across 300+ facilities (Jul 2025).
(b)/(d)	DeepFleet is a generative foundation model for fleet traffic coordination across the network, with a 10% travel-time reduction, trained on inventory-movement data in SageMaker, and it "continues to learn" (aboutamazon). It is a central shared-world-state coordinator, the closest thing to a connected brain, but for navigation only. Vulcan does tactile pick and stow and is expanding to more EU sites.
(c)	Amazon Robotics is a Thor early adopter.
(e)	Vulcan "learns from its environment and adapts with each task" (marketing).
(g)/(h)	Not sold externally; purpose-built infrastructure.

Gecko Robotics

Field	Finding
(a)	Inspection robots plus the Cantilever asset-health platform; multi-year service contracts (inspection fees, software subscription, engineering); US$1.25B valuation, US$347M raised (Jun 2025); 2023 revenue US$34.5M [secondary].
(f)	US Navy and GSA 5-year US$71M IDIQ (Mar 2026, 18 Pacific Fleet ships); ADNOC agreements (Nov 2025).
(g)	Not for explosive zones; limited vertical mobility on some assets; liability sits with Gecko as the service provider.
(h)	Inspection only; no manipulation; proprietary hardware. Relevant to ports only via ship and tank inspection.

Sanctuary AI

Field	Finding
Status	Chief executive Geordie Rose was ousted in Nov 2024 with about 30 layoffs; bridge financing, sold its Apptronik stake; new chief executive Daniel Friedmann in Jun 2026; pivot to selling Carbon physical-AI software for existing industrial robots; 99.5% success on flexible wire insertion at a Tier-1 auto supplier, 2.54 s cycle, first purchase order [single source, secondary]. Total raised about US$140M. No units shipping.
Won't do	Effectively no longer a humanoid-body company.

Field

Finding

Status

Chief executive Geordie Rose was ousted in Nov 2024 with about 30 layoffs; bridge financing, sold its Apptronik stake; new chief executive Daniel Friedmann in Jun 2026; pivot to selling Carbon physical-AI software for existing industrial robots; 99.5% success on flexible wire insertion at a Tier-1 auto supplier, 2.54 s cycle, first purchase order [single source, secondary]. Total raised about US$140M. No units shipping.

Won't do

Effectively no longer a humanoid-body company.

Generalist AI

Field	Finding
(a)/(b)	GEN-0 (late 2025) and GEN-1 (2 Apr 2026): trained on 500k+ h of wearable-device human interaction data with no robot data in the base; about 1 h of robot data per task for post-training; 99% against 64% success, about 3 times faster; "Harmonic Reasoning" at inference; end-effector-agnostic (Robot Report). Early access via a partnerships email address.
(e)	No memory or continual-learning claims.
(f)	"Thousands of robot hands shipped across new geographies", which is data collection rather than customers.

Dyna Robotics, Genesis AI, RLWRLD

Dyna: the DYNA-2 World-Action Model on 1M h of egocentric human video; the stack is DYNA-VLM for reasoning plus DYNA-System0 for control; its own DYNA-SAUR platform; deployed in hotels, restaurants and laundromats: 100 items/h laundry at 98% or better quality, 180 items/h dishware, 200 units/h factory at plus or minus 2 mm; zero-shot 87% against 46% for DYNA-1. The tagline "every deployment strengthens the fleet" comes with no disclosed mechanism (dyna.co) and is almost certainly periodic central retraining [inference].
Genesis AI: GENE-26.5 (May 2026), US$105M seed; an anatomical hand and a 1:1:1 glove data engine; a "human skill library"; a general-purpose robot "forthcoming".
RLWRLD: Korean (Seoul), not US; US$41M seed extension; reinforcement learning plus industrial data. Out of scope beyond noting that it competes for Japanese and Korean factories.

Locus Robotics

Field	Finding
(a)	RaaS; about US$35k/robot plus about US$2k/month [secondary]; Locus Array (Apr 2026, MODEX): a mobile base plus arm, autonomous picking, 30 kg totes, 3 m shelves; Nexera grasping acquisition (May 2026).
(d)	LocusONE orchestration does dynamic task assignment across Origin, Vector and Array as "a single system", which is shared work state, not shared learned skills.
(f)	7B+ picks, 150+ customers, 350+ facilities, 20 countries, 17,000 AMRs; DHL 1B picks (Feb 2026).
(g)	Wi-Fi dead spots stop robots; WMS data cleanliness is the "primary cause of implementation failure" (integrator guides); Array cannot pick jam-prone items.

Symbotic

Field	Finding
(a)/(f)	Q3 FY26 revenue US$721M, up 22%; 77 systems in deployment, 56 operational; backlog US$22.5B, of which about 15% converts in 12 months; about 85% of revenue from Walmart; BreakPack at about half of Walmart RDCs; first SymMicro store install (6-month build), potential 400-unit order; store rollout pushed to early 2028 (Investing.com transcript; Seeking Alpha).
(g)	A US$34.3M component recall; a remediation plan for cost-of-revenue controls; customer concentration.
(b)/(d)	Proprietary in-house AI agents for predictive maintenance from daily fleet data, not external LLM-token based. A closed system.

Cross-cutting answers

Who offers shared fleet memory, where robot A's experience updates robot B on the same site in near real time

Nobody sells this today. What exists, sorted by mechanism:

Central model retraining, hours to days, offline: Figure (dock-time terabyte upload, cloud aggregation, redistributed weights, explicitly not real time per ETC Journal), Tesla (FSD-style fleet data), 1X (teleoperation episodes into Redwood), Apptronik into Gemini, Dyna ("every deployment strengthens the fleet", no mechanism), Agility (over-the-air updates). None of these give robot B anything from robot A within a shift.
Shared world state and orchestration, real time but with no learned experience: Amazon DeepFleet (a network-wide traffic model, internal only), Boston Dynamics Orbit (missions, maps, Site View, an on-prem Site Hub option), Locus LocusONE, Zebra Symmetry (now Skild), Agility Arc, Certis Mozart with Field AI, Field AI Multiagent FM. These share tasks and maps, not memory of what happened or what was learned.
Shared semantic understanding across heterogeneous robots, demo: Gemini Robotics ER 2, where one cloud planner reasons over multiple robots and hands off tasks; persistence is limited to a 128k-token context; demo only (Apollo 2 plus Franka); cloud only; the terms of service exclude healthcare.
Single-robot memory, not shared: Physical Intelligence MEM, about 2 s of video short-term plus about 15 min of text long-term within one episode; Skild S1, in-context from a video prompt, no weight updates, no persistence stated; 1X, conversational memory only.

A site-scoped, on-prem, persistent short-term and long-term memory shared across a mixed fleet in near real time is open territory. The nearest competitor is a hypothetical Gemini ER 2 plus Orbit combination (cloud, US-hosted, Boston Dynamics robots only) or Skild Brain plus Symmetry (warehouse AMRs, undisclosed) [inference].

Who sells into hospitals or ports, with what results

Hospitals:

Diligent and Serve's Moxi is the only US player with scale: about 100 robots, 25+ hospitals, 1.25M+ tasks, 125k elevator rides, 575k+ staff hours claimed; annual subscription about US$200–400k per hospital [secondary]. Documented churn: MultiCare dropped 14 robots after under two years, because they needed floor escorts, got in the way and produced no real time savings (WSNA). Moxi 2.0 (Aug 2026) fixes the Wi-Fi dependence and doubles daily runtime to 18 h, which means the previous generation ran about 9 h/day.
Figure demoed pill extraction and syringe dosing with Helix 02, with no hospital customer.
Google DeepMind's ER 2 model card prohibits healthcare and safety-critical use.
Singapore context: Changi General Hospital alone runs 80+ robots; the Ministry of Health targets a 15% nurse-workload reduction via robots by 2028; no US humanoid vendor is present.

Ports:

No US humanoid or VLA player targets ports. Stretch lists Maersk as a customer, but for warehouse container and trailer unloading, not quayside. Dexterity is parcel-hub trailer loading for FedEx. Gecko inspects Navy ships and ADNOC tanks. US port automation is politically constrained by the ILWU, and only LBCT and TraPac are largely automated. PSA Tuas runs 530+ AGVs from non-US vendors. Field AI's Singapore partner Certis covers "transport hubs" for security patrols only.

What US enterprise buyers say is missing

Dexterity and hands, not locomotion. GXO chief executive Kelleher (Fortune, 31 Aug 2026): pilots with 5 humanoid vendors, "zero" in production in 2026, no ROI yet, about 2 years away; he wants "the ability to unload a truck and the dexterity to sort lipstick"; 25% warehouse turnover is the driver.
Exception handling and SKU picking. Gartner (Jan 2026): under 100 companies past proof of concept and under 20 in production for supply chain and manufacturing by 2028; models lack "dexterity, intelligence and adaptability" for SKU picking, trailer unloading and exception handling; mainstream is 10+ years away.
Integration and post-sale support. Forrester (2026): 69% of automation decision-makers are adopting or planning, but buyers now ask "how much integration is required?" and "who supports it after deployment?"; cybersecurity, liability and safety frameworks are underdeveloped.
Reliability and battery. Digit runs 4 h; humanoids generally 1 to 4 h; no vendor publishes uptime or MTBF, and Agility tracks MTBI in Arc but keeps it private.
Workflow fit and staff buy-in. The nurses' union: escorts between floors, robots in the way, no clinician input into the purchase.
Infrastructure dependence. Wi-Fi dead spots halt Locus AMRs; Moxi 1.0 needed Wi-Fi for autonomy; WMS data quality is the top implementation-failure cause.
Safety and trust. The Figure whistleblower suit; Unitree G1 root remote-code-execution CVEs (2026) and telemetry to Chinese servers; the FCC import ban of 29 Jul 2026 shows security is now a procurement-level question.
Privacy. 1X teleoperators viewing homes, with unaudited safeguards.
Vendor lock and cloud. ER 2 is cloud only; Physical Intelligence and Skild are not self-hostable; Orbit and Arc are single-vendor.
Cost and lead time. Atlas at about US$130–320k and unavailable until 2027; Figure and Tesla have no sales channel; the Thor dev kit is US$3,499 with module pricing undisclosed; Symbotic's store rollout slipped to 2028 alongside a US$34M recall.

Where US players are likely to close the gap within 18 months

Gemini ER 2 multi-robot handoff (fast iteration, but cloud and no healthcare), Skild Brain plus Symmetry (warehouse), and Moxi 2.0 (18 h, Wi-Fi-free). None of these addresses on-prem shared memory, ports, or ASEAN and Gulf support.

What this means for Tidewell

Nothing in this report changes what we have already published. Tidewell Brain has two modes: Solo runs everything a robot needs on the robot and is a first-class state rather than a fallback, and Crew connects every robot on the site network to one governed memory with core, short-term and long-term tiers, promoted into core memory only through a named human. On the body we buy joints, screws, sensors, cells and chassis, and build the rated-life sealed hand, the calibrated fingertip, the secure body bus, the certified pack and the safety file. The wheeled W1 is the volume product and the biped B1 goes only where legs are needed. We start in ports and logistics and in healthcare. What we do not claim is unchanged: no manipulation policy served over wireless, no weights updating across the fleet in real time, no 70-billion-parameter model on the robot, no whole shift held in the policy's context, no cross-site global brain.