Noble Machines Accelerates General Purpose Industrial Robot Development With NVIDIA Isaac GR00T

Factories, semiconductor facilities, foundries, and construction sites rely on people to lift heavy materials, work around complex machinery, and enter hazardous environments. These jobs are physically demanding and difficult to staff. While traditional automation can perform fixed, repetitive motions, it struggles with changing surroundings, varied objects, and multi-step workflows.

Noble Machines developed Moby, a general-purpose industrial robot, to automate demanding work that conventional robots cannot easily address. In a semiconductor warehouse, Moby may need to move cartons, trays, and specialized material carriers between unpacking stations, conveyors, and automated storage. Unlike a fixed robot, it must carry materials through existing facilities and adjust as object size, placement, and working conditions change.

Reliability is the central challenge. Moby’s hardware has to operate continuously under payload without overheating, while its whole-body controller keeps the robot and its payload stable through tight spaces, stairs, debris, and changes in posture. Its foundation model, Moby’s brain, must also keep tasks on track as lighting, box placement, and object appearance vary. Industrial customers need Moby adapted to their workflows in hours rather than after months of custom programming.

Meeting these requirements through physical testing alone is impractical. Simulation can expose Moby’s policies to thousands of conditions, but differences in friction, contact dynamics, and sensor behavior can cause policies that work virtually to struggle on physical hardware. Noble Machines needed to train and evaluate policies at scale, transfer them reliably to Moby, and run inference onboard while retaining its proprietary AI-driven, whole-body-control software.

To meet those demands, Noble Machine tapped into the NVIDIA Isaac GR00T humanoid platform for Moby’s end-to-end development. The company uses the open NVIDIA Isaac GR00T 1.7 vision-language-action model alongside Isaac Sim, Isaac Lab, Newton, NVIDIA GPUs, and Jetson Thor to connect simulation, model training, whole-body-control development, and onboard inference. Compatibility across the NVIDIA ecosystem reduces the integration work required to move from digital development to the physical robot, allowing Noble Machines to focus on Moby's hardware, whole-body control, and foundation-model capabilities.

Engineers begin in Isaac Sim, where they import Moby’s mechanical design and build digital training environments from existing simulation assets. When a customer provides a digital twin of its facility, Noble Machines can place Moby’s digital counterpart inside it. The Newton Physics Engine helps simulate contact-rich behaviors such as lifting, grasping, and moving objects, providing a controlled environment for developing tasks before running them on physical hardware. Noble Machines also co-develops humanoid actuators with Schaeffler, a motion-technology partner.

While high-fidelity simulation provides necessary scale, VR-based teleoperation offers critical real-world demonstrations. Noble Machines uses these data points to refine the open Isaac GR00T 1.7 foundation model. Concurrently, the team relies on Isaac Lab for whole-body-control training and its sim-to-real deployment workflow. Noble Machines currently monitors robotic policy performance using proprietary evaluation tools and is exploring NVIDIA Isaac Lab Arena for future validation processes. NVIDIA H100 GPUs support pre-training, while NVIDIA RTX 6000 PRO™ GPUs support post-training. Depending on scene complexity, simulations run on NVIDIA RTX 6000 PRO or NVIDIA GeForce RTX™ 5090 GPUs, enabling thousands of digital Moby instances to train in parallel. Noble Machines evaluates policies with in-house tools under controlled variations in working conditions and inputs.

Noble Machines uses an NVIDIA Jetson Thor development kit aboard Moby to process camera and other sensor data and run robotics foundation-model inference onboard. This frees Moby to make real-time decisions without relying on external communication for model inference. Separately, Noble Machines and ADLINK, an industrial edge-computing company, are developing a ruggedized, industrial-grade NVIDIA Jetson Thor system with customized connections to Moby’s sensors. Solomon, a Taiwan-based systems integrator, also helps Noble Machines adapt Moby to material-handling workflows in semiconductor facilities.

Noble's integrated development workflow shortened the path from robotics ideation to real-world use. Rather than spending years assembling and maintaining custom infrastructure, the team focused its engineering effort on the capabilities that differentiate Moby: whole-body control, foundation-model intelligence, and reliable operation in industrial environments.

This approach enabled Noble Machines to:

• Accelerate Moby's development timeline by nearly 3x, from an expected four years to 18 months.
• Ship Moby to Asia within 18 months of founding and bring two Moby3 units to a semiconductor facility for material-handling workflows.
• Achieve these milestones with a team of about 15 people, versus a conventional approach that would require at least 50 people.
• Scale robot-policy development by training thousands of digital Moby instances in parallel before deploying them to physical robots.

Noble Machines is building on Moby's work in a semiconductor facility to expand the robot’s deployment across industrial environments. The company is also exploring construction applications with a U.S. general contractor.

As it develops the next generation of Moby, Noble Machines is interested in evaluating Isaac GR00T 2 when it becomes available for future robot-policy development, NVIDIA Isaac VSLAM and NVIDIA Isaac ROS for navigation and robotics integration, and NVIDIA IGX Thor for edge computing. By extending Moby to new tasks and environments, Noble Machines plans to bring general-purpose humanoids closer to dependable industrial use.