02▲ Physical AI data and simulation infrastructure

Lightwheel 光轮智能

Integrated data, evaluation and deployment across the Physical AI lifecycle

Lightwheel has built a product system spanning the full Physical AI lifecycle around three core stages — data, evaluation and deployment. SimFoundry and EgoSuite provide scalable training data, RoboFinals delivers industrial-grade simulation evaluation, and RoboStack closes the last mile to real-robot deployment.

Lightwheel official site: humanoid robot in a simulated warehouse environment

Lightwheel's four products — SimFoundry, EgoSuite, RoboFinals and RoboStack — form an end-to-end loop from data collection and model training to simulation evaluation and real-robot deployment. The company co-builds a technology ecosystem with leading global technology companies.

01Products and solutions

Synthetic simulation data solution (SimFoundry)

Expert demonstrations are collected through teleoperation in simulation, while AutoDataGen automatically creates training tasks and diverse scenarios. Robots autonomously perform grasping, handling, navigation and assembly, generating multimodal visual, state, action, contact and reward data together with Ground Truth labels. Precision digital-twin modelling provides high physical realism, scalable domain expansion and real-time or unlimited dynamic scene generation for large-scale training.

Human video data solution (EgoSuite)

A global network for collecting high-quality, multimodal annotated human video data, supported by real-robot collection nodes, multimodal capture devices and a high-quality annotation stack. Natural human motion and interaction patterns support cross-platform robot compatibility, while data diversity improves generalisation for pre-training and imitation learning. The system supports all hardware, automates about 99% of annotation and achieves more than 98% continuous-data completeness.

Industrial-grade simulation evaluation platform (RoboFinals)

The world's first industrial-grade simulation evaluation platform for embodied intelligence, providing a scalable, reproducible and diagnostic foundation. One hundred industrial tasks deliver standardised capability measurement and large-scale comparative evaluation for advanced robot models, accelerating iteration and quantifying real capability gains. Controlled simulation variables make results reproducible, precisely locate model weaknesses and drive marginal evaluation cost toward zero.

End-to-end real-robot deployment solution (RoboStack)

An end-to-end platform connecting data collection, model training, simulation evaluation and real-robot deployment to solve the last mile. Its simulation-first strategy completes about 99% of training and evaluation in simulation and moves only about 1% of critical work into real-world alignment, shortening deployment cycles, reducing trial-and-error costs and addressing slow deployment and difficult adaptation.

02Key strengths

Deep technology collaboration with NVIDIA: co-launched the Isaac Lab-Arena industrial robot-policy evaluation benchmark, supplies synthetic data to GR00T, and is a core contributor to the SimReady asset ecosystem for Omniverse and Isaac Sim.
Co-developing a Real2Sim world-generation pipeline with Fei-Fei Li's World Labs to address evaluation at scale; jointly defining robot-evaluation industry standards with Alibaba Qwen; and integrating Siemens industrial physics models and process parameters into Lightwheel's simulation platform to bridge industrial and general embodied simulation.
LeIsaac is included in the official Hugging Face LeRobot documentation, providing industrial-grade robot simulation capabilities to more than seven million developers worldwide.
Four products close the data-evaluation-deployment loop: SimFoundry and EgoSuite scale training data, RoboFinals standardises capability measurement, and RoboStack accelerates real-robot deployment through a simulation-first strategy.

03Japan market partnerships

Lightwheel is working with Sumitomo Corporation and SCAE to bring its simulation platform to Japan and serve local OEMs and robotics companies. It has also formed a joint venture with ARTT, making Japan a priority overseas market and using ARTT's customer resources to deploy robots across diverse scenarios.

04Customers and partners

NVIDIAWorld LabsAlibaba QwenSiemensSumitomo CorporationSCAEARTT

Other companies in this category

11Data & simulation

Kaiwang Data 恺望数据

First embodied-AI data production and trading platform — ≈500,000 hours of data assets

≈500,000 hof tradeable data assets, each reused 3–4 times on average
40+ citiesin 20+ provinces, 11 long-tail scenario types, 24,000+ signed collectors

View solutions →