Lightwheel deploys NVIDIA GR00T N1.5 humanoid robots in Geely's live automotive factory using Isaac Sim simulation
Geely
Lightwheel built a simulation-first robotics platform on NVIDIA Isaac Sim, Isaac Lab, and NVIDIA Omniverse libraries to address data scarcity and the sim-to-real gap in physical AI development. For Geely, Lightwheel fine-tuned the NVIDIA Isaac GR00T N1.5 vision-language-action foundation model to the Unitree H1 humanoid robot's morphology using simulation-generated synthetic data and DexMimicGen data augmentation, then deployed it in Geely's live automotive factory where robots autonomously perform component transportation, precise part placement on inspection trays, and coordinated dual-arm manipulation for heavy components. The platform achieves a 100:1 simulated-to-real data ratio and reduced development cycles from months to weeks.
Overview
Lightwheel built a simulation-first robotics platform on NVIDIA Isaac Sim, Isaac Lab, and NVIDIA Omniverse libraries to address data scarcity and the sim-to-real gap in physical AI development. For Geely, Lightwheel fine-tuned the NVIDIA Isaac GR00T N1.5 vision-language-action foundation model to the Unitree H1 humanoid robot's morphology using simulation-generated synthetic data and DexMimicGen data augmentation, then deployed it in Geely's live automotive factory where robots autonomously perform component transportation, precise part placement on inspection trays, and coordinated dual-arm manipulation for heavy components. The platform achieves a 100:1 simulated-to-real data ratio and reduced development cycles from months to weeks.
The challenge
The embodied AI and robotics industry faces two fundamental barriers: data scarcity, since real-world data collection is slow and expensive, and the sim-to-real gap, where AI policies trained in simulation environments fail to translate to reliable performance on physical hardware — challenges that create substantial operational limitations for robotics researchers and developers deploying autonomous systems in manufacturing and other industries.
The solution
Lightwheel built the Lightwheel Simulation Platform on NVIDIA Isaac Sim, Isaac Lab, and NVIDIA Omniverse libraries. For Geely, Lightwheel fine-tuned the NVIDIA Isaac GR00T N1.5 vision-language-action foundation model to the Unitree H1 humanoid robot's morphology using simulation-generated synthetic data, teleoperation data collected via VR headsets and other devices, and DexMimicGen data augmentation across varied lighting, materials, and object placements, then deployed it in Geely's live automotive factory, where the robots autonomously perform component transportation, precise part placement on inspection trays, and coordinated dual-arm manipulation for heavy components.
Reported business value
The platform achieved a 100:1 simulated-to-real data ratio, eliminating expensive real-world data collection while maintaining physical accuracy, and reduced development cycles from months to weeks. Lightwheel successfully deployed GR00T N1.5-powered Unitree H1 humanoid robots in Geely's live automotive factory, autonomously performing tasks while maintaining balance in dynamic environments alongside human workers.
Sources
Open any source and check the claim yourself — that is the point of the register.
Other automotive entries in the register.
Volvo Trucks uses proprietary AI models to power adaptive maintenance in its Blue Service Contract
Volvo Trucks North America upgraded its Blue Service Contract to offer optimized service stops based on fuel consumption, idle time and oil samples, powered by the OEM's proprietary AI models, enabling fewer stops for 2023 model year and newer trucks and reducing operating costs for customers.
BMW Group builds plant-scale digital twins with NVIDIA Omniverse, cutting vehicle aerodynamics simulation time 30x
BMW Group is building digital twins of its production facilities using NVIDIA Omniverse libraries, letting global production planning teams collaborate in real time, optimize the layout and design of complex manufacturing systems, and develop autonomous robot and vision AI applications prior to real-world deployment. Working with Siemens, BMW is also accelerating simulation of vehicle aerodynamics: tests on NVIDIA Grace Blackwell and CUDA-X-accelerated Simcenter Star-CCM+ software showed a 30x speedup for transient aerodynamics simulations of entire vehicle geometries, while reducing energy consumption and costs.
General Motors uses AI for digital twins, battery leak detection, and weld/paint inspection in manufacturing
GM builds in-house AI solutions used across its manufacturing facilities: digital twins that simulate a running production line before it's constructed, AI/machine learning that quickly and accurately pinpoints potential leaks in battery packs, and a combination of robotics and a proprietary AI tool that inspects welds and paint coats to identify irregularities and anomalies.
F.02 Contributed to the Production of 30,000 Cars at BMW
Figure deployed its Figure 02 humanoid robot at BMW Group Plant Spartanburg for an 11-month trial performing sheet-metal loading on an active assembly line, running 10-hour shifts Monday-Friday. Over the deployment, the robot loaded 90,000+ parts, logged 1,250+ hours of runtime, and contributed to the production of 30,000+ X3 vehicles, with a target cycle time of 84 seconds and placement accuracy goal of 99% per shift.
Was this helpful?
Your feedback helps us improve our use case database