Skip to content

Research ยท Compiled from public sources

Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data

25 companies, labs and projects behind simulation and synthetic training data for robots, cars and drones: what each sells or publishes, its named customers, funding, public prices and scale, measured results of mixing synthetic with real teleoperation and human data, and what labs say about the sim-to-real gap; 60 facts, each linked to its source.

Covers
16 Nov 2022 to 28 Oct 2026
Dated by when each event occurred
Calculated
29 Oct 2026, 01:34 UTC
Sources last checked 28 Oct 2026
Where it comes from
Company press releases, product, documentation and pricing pages, arXiv papers and project pages from NVIDIA, UT Austin, UC San Diego, UC Berkeley and Toyota Research Institute, a Physical Intelligence co-founder's blog, ABI Research, TechCrunch, Bloomberg, EqualOcean, Verdict, Coverager, Tectonic, The Decoder and ADAS & Autonomous Vehicle International, each linked with a verbatim quote kept in the dataset.
Method
Revision 1. Terms defined
CC BY 4.0, citing this report

Each figure here comes as CSV and JSON files, emailed to your work address.

Need robot training data collected? We run the project: teleoperation sessions, first-person video, filmed household tasks or labeled robot video, staffed from the people who apply for this work through us, checked against your standard, starting with a paid pilot.

How we run projects

What the data shows

  1. 1

    NVIDIA is the most documented of the 25 companies, labs and projects (8 sourced facts).

    Figure 1
  2. 2

    13 statements of what 12 companies and labs sell or publish, from free simulators to world models.

    Figure 2
  3. 3

    9 funding rounds and 8 named customer lists.

    Figure 3
  4. 4

    8 measured results from 7 companies and research projects on mixing synthetic with real data.

    Figure 4
  5. 5

    The most recent fact on record (11 Mar 2026): EqualOcean reports that Lightwheel (Light Wheel Intelligence, Beijing) closed a combined Series A++ and A+++ round totalling RMB 1 billion.

    Figure 7

Figure 1

Robot simulation and synthetic data companies, labs and projects

Which companies and labs sell or publish simulators, synthetic data, digital twins and world models for training robots, cars and drones, who uses them, what they cost and what they do against human data?

Company or projectWhat it sells or publishesSourced facts
NVIDIANVIDIA NewsroomNVIDIANVIDIA's Isaac GR00T Blueprint, built on Omniverse and its Cosmos Transfer world foundation models, generates large amounts of synthetic manipulation motion data from a small number of human demonstrations.NVIDIA says Agility Robotics, Boston Dynamics, Fourier, Mentee Robotics, NEURA Robotics and XPENG Robotics simulate and train their humanoid robots in Isaac Sim and Isaac Lab, and that Skild AI uses the frameworks too.NVIDIA says its researchers used the GR00T-Dreams blueprint, which generates synthetic robot motion data with Cosmos, to build GR00T N1.5 in 36 hours instead of the nearly three months manual human data collection would have taken.NVIDIA states that its Isaac Sim robot simulator is free for internal R&D and development, including commercial R&D; a paid license is needed only to redistribute it or deliver it as a service.
Applied IntuitionApplied IntuitionApplied IntuitionApplied Intuition sells a physical AI toolchain for vehicles that uses world foundation models and neural reconstruction and rendering for high-fidelity simulation, with ML behavior agents generating scenario variation for training and validation.Applied Intuition closed a US$600 million Series F fundraise and tender offer at a US$15 billion valuation in June 2025, co-led by BlackRock-managed funds and Kleiner Perkins.n/an/a
ForetellixForetellix (PR Newswire)ForetellixForetellix's Foretify toolchain trains self-driving stacks on curated real-world driven miles augmented with generated synthetic scenarios, adding sensor simulation variations generated with NVIDIA Omniverse and Cosmos.Foretellix names Daimler Truck's Torc, Volvo, Mazda, Woven by Toyota and Nuro as customers.n/an/a
Google DeepMindGoogle DeepMindGoogle DeepMindGoogle DeepMind says it is exploring early uses of its Genie 4 interactive world model for robotics, where a Gemini model's step-by-step task plan is fed to Genie 4 as it renders the simulation.n/an/aMuJoCo, the physics simulator maintained by Google DeepMind, is free and open source.
LightwheelLightwheelLightwheelLightwheel sells SimReady simulation assets and scenes, egocentric human demonstration data (EgoSuite) and a simulation evaluation platform (RoboFinals) to robotics teams.EqualOcean reports that Lightwheel (Light Wheel Intelligence, Beijing) closed a combined Series A++ and A+++ round totalling RMB 1 billion.n/an/a
Parallel DomainParallel DomainParallel DomainParallel Domain builds replicas from real-world capture data and sells deterministic camera, lidar and radar sensor simulation, so perception systems see the same thing in simulation as on the road, in the air or on the warehouse floor.Parallel Domain raised a US$30 million Series B led by March Capital in November 2022.n/an/a
WaabiTechCrunchWaabiTechCrunch reports that Waabi trains, tests and validates its self-driving software in Waabi World, a closed-loop simulator that builds digital twins from data, simulates sensors in real time and manufactures stress-test scenarios.Waabi closed a US$750 million Series C co-led by Khosla Ventures and G2 Venture Partners in January 2026, plus a milestone-based future investment from Uber for a robotaxi partnership.Waabi's CEO Raquel Urtasun told TechCrunch the company does not need the large numbers of people and the large vehicle fleets that earlier self-driving programs needed.n/a
WayveADAS & Autonomous Vehicle InternationalWayveWayve's GAIA-3 generative driving world model has 15 billion parameters, double the size of GAIA-2, with a video tokenizer twice as large, ADAS & Autonomous Vehicle International reports.Wayve raised US$1.2 billion in a Series D in February 2026, bringing its total capital secured to US$1.5 billion at a US$8.6 billion post-money valuation, Verdict reports.ADAS & Autonomous Vehicle International reports that in early studies, testing in Wayve's GAIA-3 world model closely mirrored real-world driving results and cut the rejection rate of synthetic tests fivefold.n/a
BifrostBifrost AI (press release via Venture Capital Access Online)Bifrost AIn/aBifrost AI, a generative 3D data platform, raised an US$8 million Series A in October 2024 to meet demand for AI training data and 3D simulations.n/an/a
DexMimicGenarXivNVIDIA Research, UT Austin and UC San Diegon/aThe DexMimicGen paper reports generating 21,000 bimanual dexterous manipulation demonstrations in simulation from 60 human source demonstrations.On a real humanoid can-sorting task, the DexMimicGen authors report 90% success for a policy trained on 40 demos generated through a digital twin, against 0% for one trained on the 4 human source demos.n/a
Duality AIDuality AIDuality AI (formerly Duality Robotics)Duality AI, maker of the Falcon digital twin simulator, offers Vibe Sim, an agentic co-pilot that turns text prompts into on-demand synthetic data.n/aDuality AI quotes a Procter & Gamble scientist saying a pilot using synthetic data from its Falcon simulator cut a manufacturing ML model's error rate from 4% to 0.02%.n/a
Genesis AITechCrunchGenesis AITechCrunch reports that Genesis AI generates synthetic training data with a proprietary physics engine that it says models the physical world accurately.Genesis AI, a startup building a foundation model for all kinds of robots, emerged from stealth with a US$105 million seed round co-led by Eclipse and Khosla Ventures, TechCrunch reports.n/an/a
HillbotHillbotHillbotHillbot builds robot simulation software around the open-source ManiSkill platform and the SAPIEN simulator, pitching real-world data combined with synthetic data from simulation for robot skill learning.The paper on ManiSkill3, the platform Hillbot calls its core technology, reports simulation with rendering running 10 to 1,000 times faster than other platforms, up to more than 30,000 frames per second in benchmarked environments.n/an/a
Meta (Habitat)AI HabitatMetan/aHabitat-Sim, Meta's embodied AI simulator, renders several thousand frames per second single-threaded and over 10,000 frames per second multi-process on a single GPU.Meta's Habitat-Sim repository says the simulator is no longer actively developed or maintained by Meta's internal teams after version 0.3.4.n/a
MimicGenarXiv (CoRL 2023)NVIDIA Research and UT Austinn/aThe MimicGen paper reports generating over 50,000 robot demonstrations across 18 tasks in simulation from about 200 human demonstrations.MimicGen's authors report that its generated data compares favorably with collecting additional human demonstrations.n/a
Physical IntelligenceSergey Levine, Learning and Control (Substack)Physical Intelligencen/an/aPhysical Intelligence co-founder Sergey Levine argues on his blog that cheap stand-ins for real robot data, such as simulation or web video, work only in a few cases that match their designers' assumptions, likening them to a spork.n/a
Rendered.aiRendered.aiRendered.ain/an/an/aRendered.ai lists its synthetic data platform's Teams subscription, for generating synthetic computer vision training data on one project, at US$5,000 a month.
RoboCasaarXiv (RSS 2024)UT Austin and NVIDIA Researchn/aRoboCasa, an open kitchen simulation framework, ships 120 scenes, more than 2,500 3D objects, 100 tasks and a dataset of more than 100,000 trajectories.In RoboCasa's real-kitchen tests, co-training on its simulation data raised average success on seen objects to 24.4% from 13.6% with real data only, a 79% relative improvement.n/a
Scale AIScale AIScale AIn/aScale AI states that more than 1,000 hours of diverse demonstration data are collected and uploaded to its data platform every day.Tectonic reports that Scale AI's physical AI training data is not synthetic but collected from real-world robots, which Scale says makes it higher fidelity.n/a
Sim-and-Real Co-TrainingarXivNVIDIA, UT Austin and UC Berkeleyn/an/aA 2025 NVIDIA, UT Austin and UC Berkeley study found that co-training on simulation data raised real-world task performance by an average of 38% on a robot arm and a humanoid.n/a
Skild AIBloombergSkild AIn/aSkild AI raised about US$1.4 billion in a Series C led by SoftBank in January 2026 at a valuation above US$14 billion, Bloomberg reports.Skild AI says it pre-trains its robot foundation model on large-scale simulation and internet video, then post-trains it on targeted real-world data to deliver working solutions to customers.n/a
WaymoWaymoWaymon/aWaymo says its driver has covered nearly 200 million fully autonomous miles on public roads and billions of miles in virtual worlds.n/an/a
Genesis World (simulator)GitHub (Genesis-Embodied-AI)Genesis Embodied AI (open source)Genesis World is a robotics simulation platform that combines a multi-physics engine, a photo-realistic renderer and a cross-platform compiler behind a Python interface.n/an/an/a
Robotics simulation software marketABI ResearchABI Research (analyst estimate)ABI Research forecasts that the dedicated robotics simulation software market will reach US$1.4 billion by 2030, with design and development simulation revenue growing 21.6% a year.n/an/an/a
Toyota Research InstituteToyota Research InstituteToyota Research Instituten/aToyota Research Institute trained its Large Behavior Models on 468 hours of its own robot teleoperation data, 45 hours of teleoperation collected in simulation, 32 hours of UMI data and about 1,150 hours of internet data; it says the simulation share is small and is included so the same model can be evaluated in simulation and on real robots.n/an/a

Base: 25 companies, labs and projects, 60 facts, each linked to the page that states it (reviewed 28 Oct 2026).

Company pages, papers and filings are marked stated; what outlets and analysts report is marked reported. Results are each source's own measurement on its own tasks, so they are not compared one to one.

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Figure 2

Who sells or publishes robot simulation and synthetic data

Which companies and labs offer simulators, synthetic data generators, digital twins, sim assets and world models for robots, AVs and drones?

Company or projectAs stated
Applied IntuitionApplied IntuitionApplied Intuition sells a physical AI toolchain for vehicles that uses world foundation models and neural reconstruction and rendering for high-fidelity simulation, with ML behavior agents generating scenario variation for training and validation.as of 28 Oct 2026
Duality AIDuality AIDuality AI, maker of the Falcon digital twin simulator, offers Vibe Sim, an agentic co-pilot that turns text prompts into on-demand synthetic data.as of 28 Oct 2026
Genesis World (simulator)GitHub (Genesis-Embodied-AI)Genesis World is a robotics simulation platform that combines a multi-physics engine, a photo-realistic renderer and a cross-platform compiler behind a Python interface.as of 28 Oct 2026
Google DeepMindGoogle DeepMindGoogle DeepMind says it is exploring early uses of its Genie 4 interactive world model for robotics, where a Gemini model's step-by-step task plan is fed to Genie 4 as it renders the simulation.as of 28 Oct 2026
HillbotHillbotHillbot builds robot simulation software around the open-source ManiSkill platform and the SAPIEN simulator, pitching real-world data combined with synthetic data from simulation for robot skill learning.as of 28 Oct 2026
LightwheelLightwheelLightwheel sells SimReady simulation assets and scenes, egocentric human demonstration data (EgoSuite) and a simulation evaluation platform (RoboFinals) to robotics teams.as of 28 Oct 2026
Parallel DomainParallel DomainParallel Domain builds replicas from real-world capture data and sells deterministic camera, lidar and radar sensor simulation, so perception systems see the same thing in simulation as on the road, in the air or on the warehouse floor.as of 28 Oct 2026
WaabiTechCrunchTechCrunch reports that Waabi trains, tests and validates its self-driving software in Waabi World, a closed-loop simulator that builds digital twins from data, simulates sensors in real time and manufactures stress-test scenarios.28 Jan 2026
WayveADAS & Autonomous Vehicle InternationalWayve's GAIA-3 generative driving world model has 15 billion parameters, double the size of GAIA-2, with a video tokenizer twice as large, ADAS & Autonomous Vehicle International reports.3 Dec 2025
Genesis AITechCrunchTechCrunch reports that Genesis AI generates synthetic training data with a proprietary physics engine that it says models the physical world accurately.1 Jul 2025
ForetellixForetellix (PR Newswire)Foretellix's Foretify toolchain trains self-driving stacks on curated real-world driven miles augmented with generated synthetic scenarios, adding sensor simulation variations generated with NVIDIA Omniverse and Cosmos.21 May 2025
Google DeepMindNVIDIA NewsroomNVIDIA said Google DeepMind and NVIDIA were developing MuJoCo-Warp, expected to speed up robotics machine learning workloads by more than 70 times.18 Mar 2025
NVIDIANVIDIA NewsroomNVIDIA's Isaac GR00T Blueprint, built on Omniverse and its Cosmos Transfer world foundation models, generates large amounts of synthetic manipulation motion data from a small number of human demonstrations.18 Mar 2025

Base: 13 facts from 12 companies, labs and projects.

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Figure 3

Named customers and funding

Which robot makers, AV developers and carmakers use these simulators and world models, and how much have the companies behind them raised?

Company or projectAs stated
LightwheelEqualOceanEqualOcean reports that Lightwheel (Light Wheel Intelligence, Beijing) closed a combined Series A++ and A+++ round totalling RMB 1 billion.11 Mar 2026
LightwheelEqualOceanEqualOcean reports that Lightwheel counts NVIDIA, Google, Figure AI, 1X, ByteDance, Alibaba, Agibot, Galbot, Toyota, Bosch, BYD and Geely among its partners.11 Mar 2026
WayveVerdict (via Yahoo Finance)Wayve raised US$1.2 billion in a Series D in February 2026, bringing its total capital secured to US$1.5 billion at a US$8.6 billion post-money valuation, Verdict reports.26 Feb 2026
WaymoWaymoWaymo built its Waymo World Model simulator on Google DeepMind's Genie 3 world model, using it to simulate situations its fleet never observed.6 Feb 2026
WaabiWaabiWaabi closed a US$750 million Series C co-led by Khosla Ventures and G2 Venture Partners in January 2026, plus a milestone-based future investment from Uber for a robotaxi partnership.28 Jan 2026
Skild AIBloombergSkild AI raised about US$1.4 billion in a Series C led by SoftBank in January 2026 at a valuation above US$14 billion, Bloomberg reports.14 Jan 2026
Genesis AITechCrunchGenesis AI, a startup building a foundation model for all kinds of robots, emerged from stealth with a US$105 million seed round co-led by Eclipse and Khosla Ventures, TechCrunch reports.1 Jul 2025
Applied IntuitionApplied IntuitionApplied Intuition closed a US$600 million Series F fundraise and tender offer at a US$15 billion valuation in June 2025, co-led by BlackRock-managed funds and Kleiner Perkins.17 Jun 2025
Applied IntuitionApplied IntuitionApplied Intuition states that 18 of the top 20 global automakers and major US Department of Defense programs use its software.17 Jun 2025
ForetellixForetellix (PR Newswire)Foretellix names Daimler Truck's Torc, Volvo, Mazda, Woven by Toyota and Nuro as customers.21 May 2025
NVIDIANVIDIA NewsroomNVIDIA says Agility Robotics, Boston Dynamics, Fourier, Mentee Robotics, NEURA Robotics and XPENG Robotics simulate and train their humanoid robots in Isaac Sim and Isaac Lab, and that Skild AI uses the frameworks too.18 May 2025
NVIDIANVIDIA NewsroomNVIDIA named 1X, Agile Robots, Agility, Figure AI, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Virtual Incision, Waabi, XPENG and Uber among the first adopters of its Cosmos world foundation models when it launched them in January 2025.6 Jan 2025
BifrostBifrost AI (press release via Venture Capital Access Online)Bifrost AI, a generative 3D data platform, raised an US$8 million Series A in October 2024 to meet demand for AI training data and 3D simulations.31 Oct 2024
BifrostBifrost AI (press release via Venture Capital Access Online)Bifrost AI's CEO Charles Wong says it works with major US government organizations and heavy-industry enterprises, naming NASA JPL as a collaborator on data generation engines for Moon and Mars exploration.31 Oct 2024
ForetellixCoveragerForetellix closed an US$85 million Series C led by 83North, with NVIDIA, Woven Capital and Volvo Group VC among the investors, bringing its total raised to US$135 million, Coverager reports.5 Dec 2023
Parallel DomainParallel DomainParallel Domain raised a US$30 million Series B led by March Capital in November 2022.16 Nov 2022
Parallel DomainParallel DomainParallel Domain named Google, Continental, Woven Planet and Toyota Research Institute as customers using its synthetic data in 2022.16 Nov 2022

Base: 17 facts from 11 companies, labs and projects.

Funding stays in the currency its source gives (Lightwheel's round is in yuan).

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Figure 4

What synthetic data does against real data, measured

What did papers and companies measure when they added simulation or synthetic data to real robot data?

Company or projectResult as stated
Duality AIDuality AIDuality AI quotes a Procter & Gamble scientist saying a pilot using synthetic data from its Falcon simulator cut a manufacturing ML model's error rate from 4% to 0.02%.as of 28 Oct 2026
WayveADAS & Autonomous Vehicle InternationalADAS & Autonomous Vehicle International reports that in early studies, testing in Wayve's GAIA-3 world model closely mirrored real-world driving results and cut the rejection rate of synthetic tests fivefold.3 Dec 2025
NVIDIANVIDIA NewsroomNVIDIA says its researchers used the GR00T-Dreams blueprint, which generates synthetic robot motion data with Cosmos, to build GR00T N1.5 in 36 hours instead of the nearly three months manual human data collection would have taken.18 May 2025
Sim-and-Real Co-TrainingarXivA 2025 NVIDIA, UT Austin and UC Berkeley study found that co-training on simulation data raised real-world task performance by an average of 38% on a robot arm and a humanoid.31 Mar 2025
NVIDIANVIDIA NewsroomNVIDIA states that adding its synthetic trajectories to real data improved its GR00T N1 humanoid model's performance by 40% over training on real data alone.18 Mar 2025
DexMimicGenarXivOn a real humanoid can-sorting task, the DexMimicGen authors report 90% success for a policy trained on 40 demos generated through a digital twin, against 0% for one trained on the 4 human source demos.31 Oct 2024
RoboCasaarXiv (RSS 2024)In RoboCasa's real-kitchen tests, co-training on its simulation data raised average success on seen objects to 24.4% from 13.6% with real data only, a 79% relative improvement.4 Jun 2024
MimicGenarXiv (CoRL 2023)MimicGen's authors report that its generated data compares favorably with collecting additional human demonstrations.26 Oct 2023

Base: 8 facts from 7 companies, labs and projects.

Each result is the source's own test, often a few tasks and small real-robot trials; read it with its own setup.

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Figure 5

Scale, and how labs mix simulation with human data

How much synthetic data do these systems produce, and how do labs combine it with human teleoperation and video?

Company or projectAs stated
Meta (Habitat)AI HabitatHabitat-Sim, Meta's embodied AI simulator, renders several thousand frames per second single-threaded and over 10,000 frames per second multi-process on a single GPU.as of 28 Oct 2026
Meta (Habitat)GitHub (facebookresearch)Meta's Habitat-Sim repository says the simulator is no longer actively developed or maintained by Meta's internal teams after version 0.3.4.as of 28 Oct 2026
Scale AIScale AIScale AI states that more than 1,000 hours of diverse demonstration data are collected and uploaded to its data platform every day.as of 28 Oct 2026
WaymoWaymoWaymo says its driver has covered nearly 200 million fully autonomous miles on public roads and billions of miles in virtual worlds.6 Feb 2026
WaabiTechCrunchWaabi's CEO Raquel Urtasun told TechCrunch the company does not need the large numbers of people and the large vehicle fleets that earlier self-driving programs needed.28 Jan 2026
Scale AITectonicTectonic reports that Scale AI's physical AI training data is not synthetic but collected from real-world robots, which Scale says makes it higher fidelity.26 Sep 2025
Skild AISkild AISkild AI says it pre-trains its robot foundation model on large-scale simulation and internet video, then post-trains it on targeted real-world data to deliver working solutions to customers.29 Jul 2025
Physical IntelligenceSergey Levine, Learning and Control (Substack)Physical Intelligence co-founder Sergey Levine argues on his blog that cheap stand-ins for real robot data, such as simulation or web video, work only in a few cases that match their designers' assumptions, likening them to a spork.21 Jul 2025
Physical IntelligenceSergey Levine, Learning and Control (Substack)Sergey Levine, a Physical Intelligence co-founder, writes that adding human and even simulation data to broad, representative real-world robot data is likely to help.21 Jul 2025
Toyota Research InstituteToyota Research InstituteToyota Research Institute trained its Large Behavior Models on 468 hours of its own robot teleoperation data, 45 hours of teleoperation collected in simulation, 32 hours of UMI data and about 1,150 hours of internet data; it says the simulation share is small and is included so the same model can be evaluated in simulation and on real robots.7 Jul 2025
Sim-and-Real Co-TrainingarXivThe Sim-and-Real Co-Training authors recommend using far more simulation data than real-world data, ideally orders of magnitude more, and tuning the mix ratio.31 Mar 2025
NVIDIANVIDIA NewsroomNVIDIA says the first components of its Isaac GR00T Blueprint generated 780,000 synthetic robot manipulation trajectories, which it equates to 6,500 hours of human demonstrations, in 11 hours.18 Mar 2025
DexMimicGenarXivThe DexMimicGen paper reports generating 21,000 bimanual dexterous manipulation demonstrations in simulation from 60 human source demonstrations.31 Oct 2024
HillbotarXivThe paper on ManiSkill3, the platform Hillbot calls its core technology, reports simulation with rendering running 10 to 1,000 times faster than other platforms, up to more than 30,000 frames per second in benchmarked environments.1 Oct 2024
RoboCasaarXiv (RSS 2024)RoboCasa, an open kitchen simulation framework, ships 120 scenes, more than 2,500 3D objects, 100 tasks and a dataset of more than 100,000 trajectories.4 Jun 2024
MimicGenarXiv (CoRL 2023)The MimicGen paper reports generating over 50,000 robot demonstrations across 18 tasks in simulation from about 200 human demonstrations.26 Oct 2023

Base: 16 facts from 13 companies, labs and projects.

Scale is counted in each source's own unit (trajectories, demonstrations, hours, miles or frames per second).

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Figure 6

Prices and market size

What does robot simulation software cost where a price is public, and how big is the market?

Base: 6 facts from 4 companies, labs and projects.

Few vendors publish prices; most sell by quote. Free licences are listed as US$0.

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Figure 7

Every sourced fact

What exactly has been published about each company or project?

Company or projectKindFact
Applied IntuitionApplied IntuitionWhat it sells or publishesApplied Intuition sells a physical AI toolchain for vehicles that uses world foundation models and neural reconstruction and rendering for high-fidelity simulation, with ML behavior agents generating scenario variation for training and validation.as of 28 Oct 2026Stated by the company or an official page
Applied IntuitionApplied IntuitionFundingApplied Intuition closed a US$600 million Series F fundraise and tender offer at a US$15 billion valuation in June 2025, co-led by BlackRock-managed funds and Kleiner Perkins.17 Jun 2025Stated by the company or an official page
Applied IntuitionApplied IntuitionNamed customers and usersApplied Intuition states that 18 of the top 20 global automakers and major US Department of Defense programs use its software.17 Jun 2025Stated by the company or an official page
BifrostBifrost AI (press release via Venture Capital Access Online)FundingBifrost AI, a generative 3D data platform, raised an US$8 million Series A in October 2024 to meet demand for AI training data and 3D simulations.31 Oct 2024Stated by the company or an official page
BifrostBifrost AI (press release via Venture Capital Access Online)Named customers and usersBifrost AI's CEO Charles Wong says it works with major US government organizations and heavy-industry enterprises, naming NASA JPL as a collaborator on data generation engines for Moon and Mars exploration.31 Oct 2024Stated by the company or an official page
DexMimicGenarXivScaleThe DexMimicGen paper reports generating 21,000 bimanual dexterous manipulation demonstrations in simulation from 60 human source demonstrations.31 Oct 2024Stated by the company or an official page
DexMimicGenarXivMeasured result (synthetic vs real data)On a real humanoid can-sorting task, the DexMimicGen authors report 90% success for a policy trained on 40 demos generated through a digital twin, against 0% for one trained on the 4 human source demos.31 Oct 2024Stated by the company or an official page
Duality AIDuality AIWhat it sells or publishesDuality AI, maker of the Falcon digital twin simulator, offers Vibe Sim, an agentic co-pilot that turns text prompts into on-demand synthetic data.as of 28 Oct 2026Stated by the company or an official page
Duality AIDuality AIMeasured result (synthetic vs real data)Duality AI quotes a Procter & Gamble scientist saying a pilot using synthetic data from its Falcon simulator cut a manufacturing ML model's error rate from 4% to 0.02%.as of 28 Oct 2026Stated by the company or an official page
ForetellixForetellix (PR Newswire)What it sells or publishesForetellix's Foretify toolchain trains self-driving stacks on curated real-world driven miles augmented with generated synthetic scenarios, adding sensor simulation variations generated with NVIDIA Omniverse and Cosmos.21 May 2025Stated by the company or an official page
ForetellixForetellix (PR Newswire)Named customers and usersForetellix names Daimler Truck's Torc, Volvo, Mazda, Woven by Toyota and Nuro as customers.21 May 2025Stated by the company or an official page
ForetellixCoveragerFundingForetellix closed an US$85 million Series C led by 83North, with NVIDIA, Woven Capital and Volvo Group VC among the investors, bringing its total raised to US$135 million, Coverager reports.5 Dec 2023Reported by an outlet
Genesis AITechCrunchFundingGenesis AI, a startup building a foundation model for all kinds of robots, emerged from stealth with a US$105 million seed round co-led by Eclipse and Khosla Ventures, TechCrunch reports.1 Jul 2025Reported by an outlet
Genesis AITechCrunchWhat it sells or publishesTechCrunch reports that Genesis AI generates synthetic training data with a proprietary physics engine that it says models the physical world accurately.1 Jul 2025Reported by an outlet
Genesis World (simulator)GitHub (Genesis-Embodied-AI)What it sells or publishesGenesis World is a robotics simulation platform that combines a multi-physics engine, a photo-realistic renderer and a cross-platform compiler behind a Python interface.as of 28 Oct 2026Stated by the company or an official page
Google DeepMindMuJoCo (Google DeepMind)PriceMuJoCo, the physics simulator maintained by Google DeepMind, is free and open source.as of 28 Oct 2026Stated by the company or an official page
Google DeepMindGoogle DeepMindWhat it sells or publishesGoogle DeepMind says it is exploring early uses of its Genie 4 interactive world model for robotics, where a Gemini model's step-by-step task plan is fed to Genie 4 as it renders the simulation.as of 28 Oct 2026Stated by the company or an official page
Google DeepMindNVIDIA NewsroomWhat it sells or publishesNVIDIA said Google DeepMind and NVIDIA were developing MuJoCo-Warp, expected to speed up robotics machine learning workloads by more than 70 times.18 Mar 2025Stated by the company or an official page
HillbotHillbotWhat it sells or publishesHillbot builds robot simulation software around the open-source ManiSkill platform and the SAPIEN simulator, pitching real-world data combined with synthetic data from simulation for robot skill learning.as of 28 Oct 2026Stated by the company or an official page
HillbotarXivScaleThe paper on ManiSkill3, the platform Hillbot calls its core technology, reports simulation with rendering running 10 to 1,000 times faster than other platforms, up to more than 30,000 frames per second in benchmarked environments.1 Oct 2024Stated by the company or an official page
LightwheelLightwheelWhat it sells or publishesLightwheel sells SimReady simulation assets and scenes, egocentric human demonstration data (EgoSuite) and a simulation evaluation platform (RoboFinals) to robotics teams.as of 28 Oct 2026Stated by the company or an official page
LightwheelEqualOceanFundingEqualOcean reports that Lightwheel (Light Wheel Intelligence, Beijing) closed a combined Series A++ and A+++ round totalling RMB 1 billion.11 Mar 2026Reported by an outlet
LightwheelEqualOceanNamed customers and usersEqualOcean reports that Lightwheel counts NVIDIA, Google, Figure AI, 1X, ByteDance, Alibaba, Agibot, Galbot, Toyota, Bosch, BYD and Geely among its partners.11 Mar 2026Reported by an outlet
Meta (Habitat)AI HabitatScaleHabitat-Sim, Meta's embodied AI simulator, renders several thousand frames per second single-threaded and over 10,000 frames per second multi-process on a single GPU.as of 28 Oct 2026Stated by the company or an official page
Meta (Habitat)GitHub (facebookresearch)Stated approach to sim and real dataMeta's Habitat-Sim repository says the simulator is no longer actively developed or maintained by Meta's internal teams after version 0.3.4.as of 28 Oct 2026Stated by the company or an official page
MimicGenarXiv (CoRL 2023)ScaleThe MimicGen paper reports generating over 50,000 robot demonstrations across 18 tasks in simulation from about 200 human demonstrations.26 Oct 2023Stated by the company or an official page
MimicGenarXiv (CoRL 2023)Measured result (synthetic vs real data)MimicGen's authors report that its generated data compares favorably with collecting additional human demonstrations.26 Oct 2023Stated by the company or an official page
NVIDIANVIDIA (Isaac Sim documentation)PriceNVIDIA states that its Isaac Sim robot simulator is free for internal R&D and development, including commercial R&D; a paid license is needed only to redistribute it or deliver it as a service.as of 28 Oct 2026Stated by the company or an official page
NVIDIANVIDIA (Enterprise Licensing Guide)PriceNVIDIA AI Enterprise, the license Isaac Sim needs when it is redistributed or run as a service for third parties, lists at US$4,500 per GPU for one year (US$1,125 for education and Inception startups), or US$1 per GPU-hour on cloud marketplaces.as of 28 Oct 2026Stated by the company or an official page
NVIDIANVIDIA NewsroomMeasured result (synthetic vs real data)NVIDIA says its researchers used the GR00T-Dreams blueprint, which generates synthetic robot motion data with Cosmos, to build GR00T N1.5 in 36 hours instead of the nearly three months manual human data collection would have taken.18 May 2025Stated by the company or an official page
NVIDIANVIDIA NewsroomNamed customers and usersNVIDIA says Agility Robotics, Boston Dynamics, Fourier, Mentee Robotics, NEURA Robotics and XPENG Robotics simulate and train their humanoid robots in Isaac Sim and Isaac Lab, and that Skild AI uses the frameworks too.18 May 2025Stated by the company or an official page
NVIDIANVIDIA NewsroomScaleNVIDIA says the first components of its Isaac GR00T Blueprint generated 780,000 synthetic robot manipulation trajectories, which it equates to 6,500 hours of human demonstrations, in 11 hours.18 Mar 2025Stated by the company or an official page
NVIDIANVIDIA NewsroomMeasured result (synthetic vs real data)NVIDIA states that adding its synthetic trajectories to real data improved its GR00T N1 humanoid model's performance by 40% over training on real data alone.18 Mar 2025Stated by the company or an official page
NVIDIANVIDIA NewsroomWhat it sells or publishesNVIDIA's Isaac GR00T Blueprint, built on Omniverse and its Cosmos Transfer world foundation models, generates large amounts of synthetic manipulation motion data from a small number of human demonstrations.18 Mar 2025Stated by the company or an official page
NVIDIANVIDIA NewsroomNamed customers and usersNVIDIA named 1X, Agile Robots, Agility, Figure AI, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Virtual Incision, Waabi, XPENG and Uber among the first adopters of its Cosmos world foundation models when it launched them in January 2025.6 Jan 2025Stated by the company or an official page
Parallel DomainParallel DomainWhat it sells or publishesParallel Domain builds replicas from real-world capture data and sells deterministic camera, lidar and radar sensor simulation, so perception systems see the same thing in simulation as on the road, in the air or on the warehouse floor.as of 28 Oct 2026Stated by the company or an official page
Parallel DomainParallel DomainFundingParallel Domain raised a US$30 million Series B led by March Capital in November 2022.16 Nov 2022Stated by the company or an official page
Parallel DomainParallel DomainNamed customers and usersParallel Domain named Google, Continental, Woven Planet and Toyota Research Institute as customers using its synthetic data in 2022.16 Nov 2022Stated by the company or an official page
Physical IntelligenceSergey Levine, Learning and Control (Substack)Stated approach to sim and real dataPhysical Intelligence co-founder Sergey Levine argues on his blog that cheap stand-ins for real robot data, such as simulation or web video, work only in a few cases that match their designers' assumptions, likening them to a spork.21 Jul 2025Stated by the company or an official page
Physical IntelligenceSergey Levine, Learning and Control (Substack)Stated approach to sim and real dataSergey Levine, a Physical Intelligence co-founder, writes that adding human and even simulation data to broad, representative real-world robot data is likely to help.21 Jul 2025Stated by the company or an official page
Rendered.aiRendered.aiPriceRendered.ai lists its synthetic data platform's Teams subscription, for generating synthetic computer vision training data on one project, at US$5,000 a month.as of 28 Oct 2026Stated by the company or an official page
Rendered.aiRendered.aiPriceRendered.ai lists its Organizations subscription, with its highest compute tier and a technical account manager, at US$15,000 a month.as of 28 Oct 2026Stated by the company or an official page
RoboCasaarXiv (RSS 2024)Measured result (synthetic vs real data)In RoboCasa's real-kitchen tests, co-training on its simulation data raised average success on seen objects to 24.4% from 13.6% with real data only, a 79% relative improvement.4 Jun 2024Stated by the company or an official page
RoboCasaarXiv (RSS 2024)ScaleRoboCasa, an open kitchen simulation framework, ships 120 scenes, more than 2,500 3D objects, 100 tasks and a dataset of more than 100,000 trajectories.4 Jun 2024Stated by the company or an official page
Robotics simulation software marketABI ResearchMarket estimateABI Research forecasts that the dedicated robotics simulation software market will reach US$1.4 billion by 2030, with design and development simulation revenue growing 21.6% a year.13 Jun 2024Reported by an outlet
Scale AIScale AIScaleScale AI states that more than 1,000 hours of diverse demonstration data are collected and uploaded to its data platform every day.as of 28 Oct 2026Stated by the company or an official page
Scale AITectonicStated approach to sim and real dataTectonic reports that Scale AI's physical AI training data is not synthetic but collected from real-world robots, which Scale says makes it higher fidelity.26 Sep 2025Reported by an outlet
Sim-and-Real Co-TrainingarXivMeasured result (synthetic vs real data)A 2025 NVIDIA, UT Austin and UC Berkeley study found that co-training on simulation data raised real-world task performance by an average of 38% on a robot arm and a humanoid.31 Mar 2025Stated by the company or an official page
Sim-and-Real Co-TrainingarXivStated approach to sim and real dataThe Sim-and-Real Co-Training authors recommend using far more simulation data than real-world data, ideally orders of magnitude more, and tuning the mix ratio.31 Mar 2025Stated by the company or an official page
Skild AIBloombergFundingSkild AI raised about US$1.4 billion in a Series C led by SoftBank in January 2026 at a valuation above US$14 billion, Bloomberg reports.14 Jan 2026Reported by an outlet
Skild AISkild AIStated approach to sim and real dataSkild AI says it pre-trains its robot foundation model on large-scale simulation and internet video, then post-trains it on targeted real-world data to deliver working solutions to customers.29 Jul 2025Stated by the company or an official page
Toyota Research InstituteToyota Research InstituteScaleToyota Research Institute trained its Large Behavior Models on 468 hours of its own robot teleoperation data, 45 hours of teleoperation collected in simulation, 32 hours of UMI data and about 1,150 hours of internet data; it says the simulation share is small and is included so the same model can be evaluated in simulation and on real robots.7 Jul 2025Stated by the company or an official page
WaabiWaabiFundingWaabi closed a US$750 million Series C co-led by Khosla Ventures and G2 Venture Partners in January 2026, plus a milestone-based future investment from Uber for a robotaxi partnership.28 Jan 2026Stated by the company or an official page
WaabiTechCrunchWhat it sells or publishesTechCrunch reports that Waabi trains, tests and validates its self-driving software in Waabi World, a closed-loop simulator that builds digital twins from data, simulates sensors in real time and manufactures stress-test scenarios.28 Jan 2026Reported by an outlet
WaabiTechCrunchStated approach to sim and real dataWaabi's CEO Raquel Urtasun told TechCrunch the company does not need the large numbers of people and the large vehicle fleets that earlier self-driving programs needed.28 Jan 2026Reported by an outlet
WaymoWaymoScaleWaymo says its driver has covered nearly 200 million fully autonomous miles on public roads and billions of miles in virtual worlds.6 Feb 2026Stated by the company or an official page
WaymoWaymoNamed customers and usersWaymo built its Waymo World Model simulator on Google DeepMind's Genie 3 world model, using it to simulate situations its fleet never observed.6 Feb 2026Stated by the company or an official page
WayveVerdict (via Yahoo Finance)FundingWayve raised US$1.2 billion in a Series D in February 2026, bringing its total capital secured to US$1.5 billion at a US$8.6 billion post-money valuation, Verdict reports.26 Feb 2026Reported by an outlet
WayveADAS & Autonomous Vehicle InternationalWhat it sells or publishesWayve's GAIA-3 generative driving world model has 15 billion parameters, double the size of GAIA-2, with a video tokenizer twice as large, ADAS & Autonomous Vehicle International reports.3 Dec 2025Reported by an outlet
WayveADAS & Autonomous Vehicle InternationalMeasured result (synthetic vs real data)ADAS & Autonomous Vehicle International reports that in early studies, testing in Wayve's GAIA-3 world model closely mirrored real-world driving results and cut the rejection rate of synthetic tests fivefold.3 Dec 2025Reported by an outlet

Base: All 60 facts. "Reported" means an outlet reports it from sources the company has not confirmed.

Amounts stay in the currency and form their source gives; nothing is converted or totalled.

Credit: compiled by RobotTraining.jobs Research, Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data, checked 28 Oct 2026.

Robotics labs ยท Fleet operators ยท Data vendors ยท Research teams

Buy the data behind this report

This report sums up records we gather ourselves, day after day:

  • 1,392 robot training roles followed, history included

    Every role on 99 employer job boards, read each day since 12 Sep 2026: its pay, its place, and the first and last day it was up.

  • 471 apply clicks over 28 days

    Which robot training roles people really apply to, by kind of work, pay and place. The clicks happen here, so no one else can count them.

  • 25 sourced industry datasets

    Funding, revenue, deals, fleets, deployments and prices across the robot training industry, each record linked to the publication it comes from.

On offer

  • The raw records

    Every robot training role we track: title, employer, place, kind of work, pay as stated, and the days it was first and last open.

  • A regular feed

    New and closed roles every week or month, as file drops or through the data endpoints.

  • A custom cut

    Any figure here by country, city, kind of work, vehicle, employer or date, or your own hiring benchmarked against the market.

Need the people behind this data for your own robots? (you contract them yourself) or (we deliver the robot data).

How it was measured

Company press releases, product, documentation and pricing pages, arXiv papers and project pages from NVIDIA, UT Austin, UC San Diego, UC Berkeley and Toyota Research Institute, a Physical Intelligence co-founder's blog, ABI Research, TechCrunch, Bloomberg, EqualOcean, Verdict, Coverager, Tectonic, The Decoder and ADAS & Autonomous Vehicle International, each linked with a verbatim quote kept in the dataset. Each row carries a link to its publication and gives the number exactly as published, with no estimates or conversions. A note under every figure says what it covers.

Sim-to-real gap
The drop in performance when a robot policy trained in simulation meets the real world, because no simulator matches real physics, sensors and objects exactly.
Co-training
Training one robot model on a mixture of simulation data and real robot data at once, in a chosen ratio, rather than training in simulation and transferring.
Synthetic trajectories
Robot demonstrations generated by software (for example by replaying and adapting a few human teleoperation demos in simulation) instead of recorded one by one by people.
World model
A generative model that predicts how a scene will look and change, used to render driving or robot scenarios for training and testing (Cosmos, GAIA-3, Genie 4).
Digital twin
A simulated copy of a real place, robot or task built from real capture data, so training and tests in simulation match the real setting.

Limits of this data

  • Built only from public sources. Anything never announced or reported is absent, and private-company figures are whatever the company or the press said, unaudited.
  • Results are each paper's or company's own test, often a handful of tasks and small real-robot trials, so a 38% or 79% gain in one study does not carry over to another.
  • Few vendors publish prices: most sell by quote, so the price figure holds the free licences and the list prices that are public.
  • Customer lists are as each company or outlet states them (Lightwheel's as reported partners), and several date from 2022 to 2025; each record carries its own date.

Collection and classification of listings are explained on the methodology page. Freshness, completeness and consistency are tested daily in the data quality report.

Monthly editions

The live page shows the dataset as of its latest review, and new events join it once they are made public. Every finished month also gets its own edition, locked two days after month end, so a number quoted from it stays fixed.

  • Live page (you are here)Latest review

Quoting this report

You may quote, chart or republish these figures under CC BY 4.0, provided you link back to this page. The records change only when their sources are reviewed again, and the date of the last review is above. For a figure that stays put, quote a monthly edition, or note the date you accessed this page.

Citation

RobotTraining.jobs (2026). Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data (live data, 16 Nov 2022 to 28 Oct 2026). https://robottraining.jobs/research/robot-simulation-and-synthetic-data (accessed 29 Oct 2026)

BibTeX

@misc{robottrainingjobs_robot_simulation_and_synthetic_data_live,
  author = {{RobotTraining.jobs}},
  title = {Synthetic Data for Robotics: Robot Simulation Companies, Customers, Prices and Results vs Human Data (live data, 16 Nov 2022 to 28 Oct 2026)},
  year = {2026},
  howpublished = {\url{https://robottraining.jobs/research/robot-simulation-and-synthetic-data}},
  note = {Accessed 29 Oct 2026},
}
Files: every figure in one CSV (a row per value) or as JSON, emailed to your work address.Other open data: datasets and API

Keep reading

Browse all 91 reports