Skip to content

Physical AI

A global and Chinese landscape of robot models, machines, infrastructure, deployment and dated financing.

Updated View as Markdown

A system that generates a convincing world is not necessarily a system that can act in it. Physical AI is easier to understand when we separate representing an environment, choosing actions, collecting experience, testing behavior and operating the resulting machine.

This landscape connects global platforms with China’s embodied AI ecosystem. It includes model developers, robot makers, warehouse and service automation, and the tools that turn experiments into operating systems. Evidence was inspected October 10–11, 2026. Research releases, product demonstrations, commercial deployments and financing announcements remain distinct.

Landscape at a glance

Pan on smaller screensOpen full size ↗

Figure 1. The Physical AI learning loop, with global and Chinese examples. Author’s synthesis; AI-generated illustration.12345678

The image gives the territory at a glance; the explorer and comparisons below preserve the detail. Placement shows a product role, not market share, measured quality or a partnership. Categories overlap; arrows show conceptual connections, not a strict development order or verified integrations between companies.

Company explorer

52 companies and platforms, including 25 China-based or China-linked entries. Search a provider, task or funding stage; combine it with the category filter. Use Company base to find China-based or China-linked entries, including Flexiv’s China / U.S. footprint. Base describes the company footprint, not an exclusive sales market. Other entries are not individually country-classified; Chinese companies can serve global customers.

Read the funding column as dated evidence, not a guaranteed latest round. USD is shown as $; RMB stays in RMB, without currency conversion. Company announcements, regulatory filings, media reports and unknown stages are labeled differently. Historical disclosures do not rule out later rounds. Corporate programs inherit their parent’s status; they are not separately financed startups. Funding is not evidence of technical quality, safety, solvency or customer returns. No robot trials or ROI measurements were performed here.

Scroll the table sideways on smaller screens. Focus view gives the comparison more room on desktop.

Company / platform Category Role / buyer Product or research contribution Funding / ownership (dated) Evidence boundary / status
Google DeepMind / Gemini Robotics Robot policies Model partners and industrial developers Robotics 2 VLA, embodied reasoning and on-device models; Intrinsic adds industrial software.910 Google / Alphabet programs; not separate startup rounds.10 Model access, industrial integration and a deployed robot fleet are different offers.
Amazon Robotics / DeepFleet Fleet operation Fulfillment operations June 2025 million-robot milestone and fleet-routing model.11 Amazon business program; no separate venture stage.11 Dated operator disclosure, not universal robot intelligence or independent ROI evidence.
Tesla / Optimus Robot hardware Humanoid development and prospective operators General-purpose humanoid goal with balance, navigation, perception and interaction stacks.12 Tesla program; no separate venture stage.12 Development direction, not independently audited unattended task success.
Boston Dynamics / Atlas Robot hardware Industrial material handling Enterprise humanoid, workflow integration and fleet oversight.13 Hyundai Motor Group-owned company.14 Product and early-adopter claims; site-specific reliability and ROI still need verification.
World Labs Spatial worlds Spatial creation and simulation inputs Marble creates persistent, editable 3D worlds.1 $1B announced Feb. 18, 2026; stage unnamed. AMD agreement Sept. 28 remains conditional.1516 Creative geometry does not establish validated contact physics or robot control.
Physical Intelligence / openpi Robot policies Robot-learning researchers and integrators Public VLA models, checkpoints and adaptation examples.3 $600M reported Nov. 21, 2025; stage unspecified.17 Repository warns adaptation to a new robot may fail; not turnkey deployment.
Skild AI Robot policies Robot operators and hardware partners Robot models and data flywheel; S1 introduced Aug. 2026. Acquired Zebra’s robotics division, formerly Fetch.1819 $1.4B Series C, Jan. 14, 2026.18 Cross-body generality and customer results are company claims; not an independent audit.
Scale AI Data collection Physical-interaction dataset teams Custom collection and annotation for Physical AI.20 Meta minority investment, June 12, 2025; Scale described as independent.21 Not a Meta acquisition; data volume alone does not prove downstream gains.
Foxglove Data operations Robot developers and fleet-data teams Recording, visualization, search and curation across robot data.22 $40M Series B, Nov. 11, 2025.23 Data operations are not real-time robot control or release approval.
Roboto Data operations Robotics analytics teams Log analytics, event search and dataset workflows.24 $4.8M seed, Apr. 5, 2023. Historical; later stage unverified.25 An event match is not a causal diagnosis; customer quotes are not comparative tests.
NVIDIA / Isaac, GR00T, Cosmos Development platforms Robot developers and validation teams Isaac simulation, GR00T robot models and Cosmos world models form distinct platform layers.26272 NVIDIA platforms; not separately financed startups.262 Simulation fidelity, model transfer and deployment safety need separate evidence.
Runway Robotics Evaluation Policy-comparison researchers Learned world-model evaluation against real robot rollouts.28 Runway: $315M Series E, Feb. 10, 2026; company-level funding.29 Bounded ranking study, not calibrated failure probabilities or a safety certificate.
One Robot Evaluation Manipulation training and testing teams Task-specific simulations for textiles and box folding.30 YC company profile; round amount and stage unverified.30 Company offer, not independent proof of contact fidelity or ranking transfer.
Most Robotic / Instance launch Evaluation Robotics research and test workflows Current household-robotics profile retains an earlier success-detector and automated-test-rig launch.31 YC company profile; round amount and stage unverified.31 Do not assume the original product or entire rollout/reset rig is generally available.
Unitree Robot hardware Research platforms and robot integrators Humanoid and quadruped platforms, including G1 and H-series hardware.32 Listed on Shanghai STAR Market Aug. 19, 2026.33 Hardware sales and athletic demonstrations do not establish autonomous task reliability.
AGIBOT Robot hardware Industrial, service and research buyers A2, X2 and G2 portfolio; AGIBOT World adds data and GO-1 model resources.345 Private company; round amount and stage not verified in inspected sources. Shanghai humanoid company, not the similarly named medical-robotics company. Deployment claims remain attributed.
Galbot Robot policies Manipulation and industrial integration Embodied models, manipulation research and G1 applications connect learning to robot tasks.3536 >$300M new round, Dec. 19, 2025; stage unnamed. Historical disclosure.36 Separate from Galaxea. Generalization and deployment scale are company claims.
UBTECH Robot hardware Industrial humanoid operators Walker S-series humanoids; 2025 annual report describes Walker S2 commercialization and delivery.8 Listed HKEX 9880; 2025 annual report published Apr. 14, 2026.8 Issuer disclosure is stronger than a demo, but not an independent task-reliability study.
LimX Dynamics Robot hardware Embodied developers and integrators Oli humanoid, modular TRON2 platform and COSA embodied software stack.37 $200M Series B, Feb. 2, 2026.37 Platform flexibility is not evidence of transfer to every task or embodiment.
X Square Robot policies Embodied model and robot customers Full-stack embodied models, robot hardware and proprietary data pipelines.438 Four consecutive rounds culminating in Series C, closed June 29, 2026; single-round amount not stated.38 Use the newer financing release, not the older English timeline; demonstrations are not a fleet audit.
Galaxea AI Robot policies Robot researchers and integrators G0.5 VLA, open-world data and deployment entrypoints for R1 Lite / Pro.39 Round amount and current stage not verified in inspected primary sources. Distinct from Galbot. Current community-licensed materials require separate permission for commercial use.39
Spirit AI Robot policies Industrial manipulation teams VLA development using human and robot data, with Moz1 industrial applications.40 $280M funding announced Feb. 25, 2026; stage unnamed.40 Factory performance figures are company reports, not cross-task independent validation.
Robotera Robot hardware Logistics and humanoid integration Humanoid systems; funding release describes M7 parcel-handling applications with SF.41 >$200M new round, May 8, 2026; stage unnamed.41 Named customer use is not proof of general autonomy or fleet-wide reliability.
EngineAI Robot hardware Humanoid developers and operators PM01, SE01 and T800 hardware; company reports first T800 production in May 2026.42 $200M Series B, Apr. 10, 2026.42 Production capability and locomotion displays are not measured unattended commercial task performance.
Fourier Robot hardware Social interaction and embodied research GR-3 Care-bot humanoid with visual, auditory and tactile interaction systems.43 Private company; current round amount and stage unverified. Care-oriented positioning does not establish clinical efficacy, safe autonomy or deployment scale.
Astribot Robot hardware Manipulation research and service integration S1 and T1 robot systems; public research updates include learned manipulation.44 Private company; current round amount and stage unverified. Demonstrated skills need testing on the buyer’s tasks, failure cases and supervision limits.
XPENG / IRON Robot hardware Prospective humanoid operators IRON production line commissioned Sept. 8, 2026; external deliveries described as a 2027 plan.45 >$900M robotics-business share purchase agreements announced Aug. 24, 2026; not a separately verified closing.46 Commissioned production line ≠ completed customer rollout. Parent is listed; subsidiary financing is separate.
Pudu Robotics Service robotics Hospitality, cleaning and industrial delivery Commercial delivery, cleaning, industrial transport and embodied robot product lines.47 Nearly $150M new round, Apr. 23, 2026; stage unnamed.47 A useful bounded-workflow comparator, not evidence that all service tasks are automated.
KEENON Robotics Service robotics Hospitality, facilities and delivery buyers DINERBOT delivery, CLEANBOT cleaning and additional robot product lines.48 Private company; inspected undated timeline does not establish a dated current round.48 Commercial product categories should not be equated with general-purpose humanoid autonomy.
Geek+ Warehouse automation Warehouse and supply-chain operators Mobile robotics for storage, picking and sorting workflows.49 Listed HKEX 2590, July 9, 2025.50 Compare complete warehouse integration, throughput and support, not a generalist-model leaderboard.
Hai Robotics Warehouse automation Warehouse storage and fulfillment HaiPick and Climb systems combine container handling with warehouse workflows.51 >$100M Series D+, June 15, 2022. Historical; later financing / listing unverified.52 Deployment accounts are company disclosures; current IPO completion is not assumed.
Flexiv Industrial automation Industrial manipulation and integration Force-sensing adaptive robots and AI-enabled industrial applications.53 Invus strategic investment, Mar. 18, 2026; amount and named stage undisclosed.53 China / U.S. footprint. Adaptive industrial arms are not a general-purpose humanoid product.
Dobot Industrial automation Collaborative automation buyers Collaborative robot arms and manufacturing applications; humanoid products are a separate line.54 Listed HKEX 2432, Dec. 23, 2024.54 An established cobot business does not by itself validate newer humanoid capabilities.
Alibaba DAMO / RynnBrain Robot policies Embodied-model researchers Released embodied understanding and grounding resources for robot-oriented perception and reasoning.55 Alibaba research program; no separate startup round.55 Understanding / grounding is not by itself a complete closed-loop control or commercial robot system.
Ant Group / Robbyant Robot policies Embodied-model researchers and integrators LingBot models cover robot action and spatial / world understanding; public code and weights.5657 Company under Ant Group; separate round amount and stage not established.56 Model-family breadth and released artifacts are not an independent field-reliability audit.
ByteDance Seed / Robotics Robot policies Robot-learning researchers GR-3 and GR-RL robotics research in the Seed model catalog.58 ByteDance research program; no separate startup round.58 Research models are not an orderable consumer robot or proven universal task coverage.
D-Robotics Development platforms Robot OEMs and developers Sunrise computing and RDK hardware / software development ecosystem.59 $400M Series C, Sept. 17, 2026.59 Computing infrastructure, not a humanoid maker; do not substitute Horizon’s listing status.
DEEP Robotics Robot hardware Industrial inspection and embodied developers Industrial quadrupeds, LYNX systems and DR humanoid development.60 >RMB500M Series C, Dec. 9, 2025.60 Industrial locomotion / inspection is a different task distribution from household manipulation.
Leju / Kuavo Robot hardware Humanoid researchers and industrial teams Kuavo hardware, Kuavo-VLA and LeTools for collection, training and deployment.61 Round amount and current stage not verified in inspected sources. Public developer tools and benchmark results do not establish general production reliability.
Figure Robot hardware Industrial and prospective home operators Figure humanoids and Helix model development.62 >$1B committed capital in Series C, Sept. 16, 2025.63 Committed capital ≠ received proceeds; demonstrations are not unattended household readiness.
1X / NEO Robot hardware Early-access household buyers NEO home robot with basic autonomy and scheduled Expert Mode for unfamiliar chores.64 $100M Series B, Jan. 13, 2024. Historical; later stage unverified.65 Remote human help is part of the operating model, not evidence of complete autonomous coverage.
Apptronik / Apollo Robot hardware Manufacturing and logistics integration Apollo humanoid development with industrial partners.66 $520M Series A-X extension, Feb. 11, 2026; total Series A >$935M.66 Partner participation and funding do not independently establish scaled task reliability.
Agility Robotics / Digit Robot hardware Warehouse and logistics operators Digit 5 robot and Nerve fleet / workflow platform.67 Proposed Churchill Capital XI combination announced June 24, 2026; completion not established here.68 Warehouse workflow evidence is not general household capability; proposed public transaction ≠ listing.
Sanctuary AI Industrial automation Industrial dexterity customers Robotic hands and Physical AI for industrial manipulation.69 Zeon strategic investment, May 26, 2026; amount and stage undisclosed.70 Canadian dexterity platform; material and pilot partnerships are not audited operating economics.
Dexterity / Mech Industrial automation Logistics and manufacturing operators Purpose-built dual-arm machine for loading, unloading and palletizing.71 $95M reported Mar. 11, 2025; stage unnamed.72 Judge shift-level throughput, exception handling and support rather than resemblance to a human.
FieldAI Robot policies Industrial autonomous robot operators Field Foundation Models for risk-aware autonomy across robot platforms.73 $405M across two rounds, announced Aug. 20, 2025; not a single named round.73 Body-independent autonomy and scale remain company claims within specified deployments.
ANYbotics Industrial automation Asset inspection and facilities ANYmal inspection robots and related deployment software.7475 $60M additional funding, Dec. 12, 2024; historical. Later strategic investment listed; terms not verified.7475 Inspection deployment is not general manipulation; older disclosed amount is not labeled latest.
NEURA Robotics Robot hardware Cognitive robotics and industrial integration Cognitive robot portfolio and Neuraverse platform.76 Up to $1.4B Series C, June 10, 2026.76 Retain “up to”; large funding and production ambitions do not establish delivered autonomous capacity.
DYNA Robotics Robot policies Repeated commercial manipulation tasks Robot models and task deployments built around real-world learning loops.7778 $120M Series A, Sept. 2025.78 Task-specific operating accounts need independent evidence of supervision, recovery and economics.
Generalist Robot policies Robot-model partners Embodied model and data development, rather than manufacturing the robot body.79 Aug. 24, 2026 Form D reports about $198.2M sold; $208.2M offered. Stage not specified.80 Offering target ≠ capital already sold; public demonstrations do not prove universal transfer.
Hugging Face / LeRobot Development platforms Open robot-learning community Policies, datasets, hardware integrations and a shared learning toolchain.6 Company-level Series D: $235M reported Aug. 24, 2023. Historical; later status unverified.81 A community ecosystem, not a robot fleet; inspect each artifact’s license and safety requirements.
SignalFlag / formerly ReSim Evaluation Physical AI testing and release teams Test orchestration across existing simulation, replay, hardware and field logs.7 Current funding round amount and stage unverified. ReSim redirects to SignalFlag. It organizes evidence; it does not supply its own simulator.7

What China adds to the landscape

China is not one humanoid category. The useful picture has three overlapping layers: learning systems, robot bodies and commercial workflows. That grouping is my interpretation of the inspected products, not a market-share estimate.

Layer Selected examples What to compare
Models, data and developer stacks X Square, Spirit AI, Galbot, Galaxea, Robbyant, Alibaba DAMO, ByteDance Seed, AGIBOT World, D-Robotics.4403539565558559 Action interface, data coverage, adaptation, license and supported hardware.
Bodies and integrated systems Unitree, AGIBOT, UBTECH, LimX, EngineAI, Robotera, Fourier, Astribot, XPENG, DEEP Robotics and Leju.323483742414344456061 Task success, supervision, recovery, maintenance and integration cost.
Commercial workflow comparators Geek+, Hai Robotics, Pudu, KEENON, Flexiv and Dobot.495147485354 Completed work per shift, interventions, installed workflow and service support.

Models and bodies are becoming more closely connected

AGIBOT publishes model and data resources alongside its robot portfolio. Galaxea exposes VLA deployment entrypoints for its R1 hardware. Leju’s LeTools connects collection, training and deployment on Kuavo. These are concrete examples of a hardware and learning loop, not proof that vertical integration is always the better business model.3453961

My interpretation: integration can make data collection and testing easier to coordinate, but it can also bind a model to a particular sensor setup, action interface or robot body. Ask what transfers without new demonstrations, what requires adaptation and who owns the failure when components are combined.

The established businesses matter as much as the humanoid demos

Warehouse storage, cleaning, delivery and adaptive industrial arms are different problems from a general-purpose household robot. Geek+, Hai Robotics, Pudu, KEENON, Flexiv and Dobot provide useful comparisons because their products are organized around identifiable workflows.495147485354

That does not independently prove their economics. It does give a more useful question than “Which robot looks most capable?”: What work is completed, under what operating constraints, and with how much human help?

Open resources do not mean unrestricted deployment

AGIBOT World, GalaxeaVLA, LingBot and LeRobot expose different combinations of data, models, code and tools.539576 Galaxea’s current repository explicitly distinguishes historical Apache-licensed materials from newer community-licensed materials requiring separate permission for commercial use.39 Inspect the actual artifact and intended use, not only an “open” label. Availability is also not evidence of safe robot control.

What the global coverage audit added

The earlier map underrepresented robot makers and bounded industrial products. Figure, 1X, Apptronik and Agility add humanoid operating models. Sanctuary, Dexterity and ANYbotics add dexterity, logistics and inspection. FieldAI, Generalist and DYNA broaden the model and commercial-task layer; NEURA adds a cognitive robotics platform.6264666769717473797776

Two infrastructure gaps also matter. LeRobot adds the open learning ecosystem. SignalFlag, formerly ReSim, adds test orchestration across tools a team already uses. It is not another simulator. NVIDIA’s entry now separates world models, robot policies and simulation rather than treating Isaac Sim as the entire platform.6722726

This remains a selected landscape, not a census. Autonomous-driving fleets, drones, surgical robotics, conventional industrial robot catalogs and component-only suppliers would each need a separate study. Research institutes may shape the ecosystem, but they are not added to the company count.

How to read financing without reading too much into it

The distinctions below change the meaning of a headline. They are not cosmetic.

Disclosure Example What it does not establish
Named round LimX: Series B; D-Robotics: Series C.3759 No later round, safe autonomy or positive unit economics.
Committed capital Figure’s Series C announcement says committed capital.63 That every dollar has already been received.
Agreements signed XPENG robotics financing announcement describes share purchase agreements.46 Independently confirmed closing or customer delivery.
Securities sold Generalist’s Form D separates about $198.2M sold from $208.2M offered.80 A named Series stage or permission to add the offering target to proceeds.
Public status Unitree, UBTECH, Geek+ and Dobot have listing evidence.3385054 General technical leadership or the same maturity across all product lines.
Proposed transaction Agility’s public combination; World Labs’ AMD agreement.6816 A completed listing or acquisition.
Undisclosed / unverified Several private companies do not have a verified dated round in the inspected sources. Zero funding, poor quality or proof that no round occurred.

The funding column helps explain ownership and resources. The deployment column helps explain what has actually been offered, released or reported. Neither is a substitute for the other.

A value chain with a feedback loop

Figure 1 is a functional map, not a set of isolated markets. A provider may span several roles. Deployment produces observations and failures that can inform data selection, training and further tests; that feedback is my synthesis of the inspected model and data-platform accounts.182224

Lens What it does A useful test
Represent the world Construct or predict environments for creation, training and tests. What geometry, physics and action responses are actually validated?
Choose actions Map observations and goals to robot behavior. Which robots, tasks and environments does the policy work on?
Collect experience Build and annotate the demonstrations and interactions needed for learning. Does more data add useful coverage or repeat familiar situations?
Organize and diagnose Find events, investigate behavior and curate datasets across recordings. Can an observed failure become a reproducible test and useful data?
Validate and operate Combine policy tests, controlled hardware trials and monitored deployment. What evidence justifies release, and what stops an unsafe run?

Feedback: operation → recorded behavior → investigation and data selection → updated policy → fresh evaluation. The loop is not a guarantee of improvement.

Four boundaries worth keeping clear

A platform, a robot and an operator are different businesses

Google DeepMind’s VLA and embodied-reasoning models supply distinct capabilities. NVIDIA supplies simulation infrastructure, robot policies and world models as distinct layers.272 Boston Dynamics markets Atlas for industrial workflows; Tesla’s page describes the Optimus development goal. Amazon’s June 2025 announcement reports its million-robot milestone and a fleet-coordination model.926131211

My interpretation: these are complementary roles, not interchangeable “robotics companies.” Compare integration and model access for a platform, task performance and serviceability for hardware, and throughput, intervention and cost for an operator. A development goal, an early-adopter invitation and a deployed-fleet disclosure are different evidence states.

A world model and a policy have different jobs

A world model represents or predicts an environment. A policy chooses what the robot should do. In Runway’s evaluation experiment, each policy chooses actions in response to observations generated by the model. openpi exposes robot observations and action outputs rather than only generating a visual scene.283

“World model” therefore does not identify a complete product category. A navigable creative scene, a physically based simulator and an action-conditioned prediction used in a control loop have different jobs. World Labs’ Marble is a spatial creation product; that alone does not establish validated contact physics or robot-control performance.1

The company also announced a definitive agreement to join AMD on September 28, 2026. The announcement expects closing by the end of 2026, subject to conditions. An announced agreement is not a completed transaction.16

Generality is conditional

Skild describes learning across robot bodies. Physical Intelligence releases models that practitioners can adapt. Compare the supported body, observations, action interface, task distribution and adaptation procedure, not only the word “generalist.” The openpi warning that adaptation to a new robot may fail is useful counterevidence to the assumption that a checkpoint works everywhere.183

For an operator, the practical question is narrower: what does this system do reliably on my hardware, in my environment, with my acceptable level of supervision? A demonstration on another body is a lead, not the acceptance test.

A good ranking is not a safety certificate

Runway’s February 2026 study evaluated eight policies, using 1,450 simulated rollouts, on RoboArena manipulation tasks with a Franka Panda arm. It reports 0.95 Pearson correlation between simulated and real policy scores. The authors distinguish relative ranking from predicting absolute success rates.28

What the reported result supports What it does not establish
Comparing candidates within the studied setup General transfer to every robot or task
A promising way to reduce some hardware evaluation work Calibrated probabilities of failure
Agreement between simulated and real policy scores Safe unattended deployment

This interpretation follows the study’s stated scope and limitations, not a claim of independent reproduction.28 Correlation alone also cannot tell an operator which rare failure modes were missed.

Follow one failure through the value chain

Consider a robot that occasionally drops an object. The example below is an illustrative workflow, not a reported incident or a guarantee that a provider solves it.

Step Evidence needed Role in the landscape
Observe Synchronized video, state, actions and deployment versions Recording and data operations
Diagnose Repeatable conditions and competing explanations Search, visualization and analytics
Select data Cases that add coverage, with evaluation held separate Collection, annotation and curation
Compare policies Reproducible tasks, interventions and failure severity Simulation and policy evaluation
Release Hardware evidence, operating limits and a recovery procedure Integration and operation

The product opportunity is to connect these steps. Surfacing an anomaly is useful, but the operator still needs a reproducible case, a justified change and evidence that it helped without introducing a regression. That connection is my synthesis of the data, policy and evaluation roles, not a comparative measurement of the listed companies.2224328

More data is not automatically the right next step. The failure could involve perception, timing, calibration, grasp selection or changing objects. A useful workflow helps choose between explanations before collecting a larger dataset or training another model. Those are diagnostic questions, not asserted causes of an actual failure.

Evaluation is becoming a product category

SignalFlag orchestrates tests across existing simulation, replay and hardware evidence; it does not create its own simulator.7 Runway and One Robot approach evaluation through learned environments. One Robot describes task-specific simulations for hard manipulation such as textiles and box folding. The retained Instance launch starts with detecting whether a real rollout succeeded, with a broader automated judgment and reset rig as its ambition. These address distinct bottlenecks: candidate comparison, task fidelity, human judgment and hardware availability.283031

Evaluation product What a buyer should ask
Learned simulation Does it preserve the decisions that matter on real hardware?
Success detector What are false-success rates, uncertain cases and transfer limits?
Test orchestration Can a failure be reproduced across simulator, logs, software and hardware versions?
Automated test rig Are resets correct, trials reproducible and failures safely handled?
Release evidence Which conditions were tested, which remain excluded and who approves operation?

These are proposed acceptance criteria. A launch description does not establish independent performance. Most Robotic’s current YC profile also differs from the retained Instance launch, so the table does not assume that the original product or full rig is currently generally available.31

Compare economics at the operator’s boundary

An API price or training benchmark captures only part of the cost. For a robot operator, compare completed work under a stated task distribution, including integration, downtime, interventions, resets, recovery and maintenance. A successful trial with a nearby researcher is not the same operating model as reliable work across a fleet.

For a data platform, measure the time from a recorded failure to a useful diagnosis or dataset. For an evaluation tool, measure whether it selects better candidates and avoids expensive wrong decisions, not just how many cheap simulations it runs. For a collection service, measure useful coverage and downstream gains rather than hours of data alone.

The key distinction is better research throughput versus better operating performance. Faster experiments can be valuable without proving that a robot is ready for unattended use. Both outcomes need evidence, but not the same evidence. These are proposed product measures, not ROI results collected here.

Open questions

  • How much useful coverage comes from each additional demonstration or deployment?
  • When does simulated evidence change a hardware testing decision correctly?
  • Can a team reproduce a failure across data, policy, hardware and environment versions?
  • Who owns integration, supervision, recovery and the final release decision?
  • Which part of the improvement loop is expensive enough for a customer to pay to shorten it?

Return to AI Landscapes for the product view. Use Multimodal integration for the language and vision foundations; it does not replace a robotics study.

References

  1. World Labs. Spatial intelligence and Marble. Checked October 10, 2026. Used for product descriptions, not undocumented architecture or robot-control claims. ↩ ↩2 ↩3

  2. NVIDIA. Cosmos world foundation models. Inspected October 11, 2026. ↩ ↩2 ↩3 ↩4 ↩5

  3. Physical Intelligence. openpi repository, README, model types, checkpoints and adaptation caveats. Inspected October 10, 2026; the inspected public repository is not assumed to contain every model the company has developed. ↩ ↩2 ↩3 ↩4 ↩5

  4. X Square. Company overview. Inspected October 11, 2026. ↩ ↩2 ↩3

  5. AGIBOT / OpenDriveLab. AGIBOT World repository. Inspected October 11, 2026. ↩ ↩2 ↩3 ↩4 ↩5

  6. Hugging Face. LeRobot documentation. Inspected October 11, 2026. ↩ ↩2 ↩3 ↩4

  7. SignalFlag. Physical AI test and evaluation platform. Inspected October 11, 2026. ↩ ↩2 ↩3 ↩4 ↩5

  8. UBTECH / HKEX. 2025 annual report, 2026-04-14. ↩ ↩2 ↩3 ↩4 ↩5

  9. Google DeepMind. Gemini Robotics model family. Undated living page; inspected October 10, 2026. ↩ ↩2

  10. Google. Intrinsic joins Google, 2026-02-25. ↩ ↩2

  11. Scott Dresser / Amazon. Millionth robot and DeepFleet announcement. June 30, 2025 announcement; inspected October 10, 2026. Fleet counts and efficiency figures are dated company disclosures. ↩ ↩2 ↩3

  12. Tesla. AI and robotics. Undated company page; inspected October 10, 2026. Optimus is presented as a development goal, not audited commercial autonomy. ↩ ↩2 ↩3

  13. Boston Dynamics. Atlas enterprise humanoid. Undated product page; inspected October 10, 2026. Early-adopter language and product claims do not establish fleet scale or customer ROI. ↩ ↩2

  14. Boston Dynamics. Funding and ownership FAQ. Inspected October 11, 2026. Majority interest acquired by Hyundai Motor Group in June 2021. ↩

  15. World Labs. Funding announcement, 2026-02-18. ↩

  16. World Labs. World Labs is Joining AMD, September 28, 2026. Checked October 10, 2026; announced agreement, with closing conditions still stated in the source. ↩ ↩2 ↩3

  17. Axios. Physical Intelligence funding, November 21, 2025. ↩

  18. Skild AI. Announcing Series C, January 14, 2026, “The Skild Data Flywheel.” Inspected October 10, 2026; company claims remain attributed. ↩ ↩2 ↩3 ↩4

  19. Skild AI. Company research and product updates. Inspected October 11, 2026. ↩

  20. Ben Levin, Scale AI. Expanding Our Data Engine for Physical AI, September 24, 2025. Inspected October 10, 2026; used for collection and annotation offerings, not estimated data requirements. ↩

  21. Scale AI. Next phase of company evolution, 2025-06-12. ↩

  22. Foxglove. Robot data lifecycle and platform. Checked October 10, 2026. ↩ ↩2 ↩3

  23. Foxglove. Series B announcement, 2025-11-11. ↩

  24. Roboto. Robotics analytics platform. Checked October 10, 2026; customer quotes are not independent comparative tests. ↩ ↩2 ↩3

  25. Roboto. Seed funding announcement, 2023-04-05. ↩

  26. NVIDIA. Isaac Sim. Checked October 10, 2026, simulation, synthetic data and testing descriptions. ↩ ↩2 ↩3 ↩4

  27. NVIDIA. Isaac GR00T repository. Inspected October 11, 2026. ↩ ↩2 ↩3

  28. Andy Chen, Lucas Eager Leavitt, Rik Heijdens and Robin Kahlow, Runway Robotics. Accelerating Robot Policy Evaluation with General World Models, February 27, 2026, methodology, results and limitations. Inspected October 10, 2026. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  29. Runway. Series E funding, 2026-02-10. ↩

  30. One Robot. Company profile and world-model simulation account. Checked October 10, 2026. Company-reported training and evaluation capabilities, not independent benchmarks. ↩ ↩2 ↩3

  31. Most Robotic. Company profile and retained Instance launch. Checked October 10, 2026. The launch explicitly starts with success detection; the broader automated rig is an ambition, not assumed general availability. ↩ ↩2 ↩3 ↩4

  32. Unitree. Robot products and news. Inspected October 11, 2026. ↩ ↩2

  33. Xinhua / Shanghai Stock Exchange. Unitree listing report, 2026-08-27. ↩ ↩2

  34. AGIBOT. MWC 2026 robot portfolio, 2026-03-02. ↩ ↩2 ↩3

  35. Galbot. Company and embodied model overview. Inspected October 11, 2026. ↩ ↩2

  36. Galbot, via PR Newswire. New funding announcement, 2025-12-19. ↩ ↩2

  37. LimX Dynamics. Series B announcement, 2026-02-02. ↩ ↩2 ↩3 ↩4

  38. X Square, via PR Newswire. Four rounds culminating in Series C, 2026-06-29. ↩ ↩2

  39. Galaxea. GalaxeaVLA and open-world resources. Inspected October 11, 2026. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  40. Spirit AI. Funding and embodied learning announcement, 2026-02-25. ↩ ↩2 ↩3

  41. Robotera, via PR Newswire. New funding round, 2026-05-08. ↩ ↩2 ↩3

  42. EngineAI. Product news and financing timeline. Inspected October 11, 2026. ↩ ↩2 ↩3

  43. Fourier. GR-3 product documentation. Inspected October 11, 2026. ↩ ↩2

  44. Astribot. Robot products and research. Inspected October 11, 2026. ↩ ↩2

  45. XPENG. IRON production line commissioned, 2026-09-08. ↩ ↩2

  46. XPENG. Robotics financing agreements, 2026-08-24. ↩ ↩2

  47. Pudu Robotics. New funding and portfolio, 2026-04-23. ↩ ↩2 ↩3 ↩4

  48. KEENON. Company and service robot portfolio. Inspected October 11, 2026. ↩ ↩2 ↩3 ↩4

  49. Geek+. Warehouse robotics platform. Inspected October 11, 2026. ↩ ↩2 ↩3

  50. Geek+. HKEX listing announcement, 2025-07-09. ↩ ↩2

  51. Hai Robotics. Warehouse products and news. Inspected October 11, 2026. ↩ ↩2 ↩3

  52. Hai Robotics. Series D+ financing, 2022-06-15. ↩

  53. Flexiv. Strategic investment from Invus, 2026-03-18. ↩ ↩2 ↩3 ↩4

  54. Dobot. HKEX listing and collaborative robot portfolio, 2024-12-23. ↩ ↩2 ↩3 ↩4 ↩5

  55. Alibaba DAMO Academy. RynnBrain repository. Inspected October 11, 2026. ↩ ↩2 ↩3

  56. Robbyant. Official organization profile. Inspected October 11, 2026. ↩ ↩2 ↩3

  57. Robbyant. LingBot-VLA repository. Inspected October 11, 2026. ↩ ↩2

  58. ByteDance Seed. Model catalog. Inspected October 11, 2026. ↩ ↩2 ↩3

  59. D-Robotics. Series C and robotics computing stack, 2026-09-17. ↩ ↩2 ↩3 ↩4

  60. DEEP Robotics. Series C and embodied robotics portfolio, 2025-12-09. ↩ ↩2 ↩3

  61. Leju. LeTools and Kuavo robot learning. Inspected October 11, 2026. ↩ ↩2 ↩3

  62. Figure. Figure and Helix products. Inspected October 11, 2026. ↩ ↩2

  63. Figure. Series C announcement, 2025-09-16. ↩ ↩2

  64. 1X. NEO early access. Inspected October 11, 2026. ↩ ↩2

  65. 1X. Series B announcement, 2024-01-13. ↩

  66. Apptronik. Series A extension and Apollo, 2026-02-11. ↩ ↩2 ↩3

  67. Agility Robotics. Digit and Nerve. Inspected October 11, 2026. ↩ ↩2

  68. Agility Robotics. Investor relations and proposed combination. Inspected October 11, 2026. ↩ ↩2

  69. Sanctuary AI. Industrial dexterity platform. Inspected October 11, 2026. ↩ ↩2

  70. Zeon Ventures. Investment in Sanctuary AI, 2026-05-26. ↩

  71. Dexterity. Mech industrial robotics. Inspected October 11, 2026. ↩ ↩2

  72. TechCrunch. Dexterity financing, March 11, 2025. ↩

  73. FieldAI. Funding and autonomy platform, 2025-08-20. ↩ ↩2 ↩3

  74. ANYbotics. Funding and industrial inspection, 2024-12-12. ↩ ↩2 ↩3

  75. ANYbotics. Company news and product links. Inspected October 11, 2026. ↩ ↩2

  76. NEURA Robotics. Series C announcement, 2026-06-10. ↩ ↩2 ↩3

  77. DYNA Robotics. Commercial robot platform. Inspected October 11, 2026. ↩ ↩2

  78. DYNA Robotics. Series A, 2025-09. ↩ ↩2

  79. Generalist. Next phase of Physical AI, 2026-06-04. ↩ ↩2

  80. Generalist / SEC. Form D, offering and sales, 2026-08-24. ↩ ↩2

  81. Axios. Hugging Face financing, August 24, 2023. ↩

Navigation

Type to search…

↑↓ navigate↵ selectEsc close