NVIDIA says Noble Machines uses VR teleoperation and thousands of simulated Moby robots, and can place a digital Moby inside a customer’s facility twin when one is provided. Noble says it delivered its first robot within 18 months. GR00T 1.7 was publicly released after Noble announced that milestone, while site-level production results remain undisclosed.
NVIDIA has detailed how Noble Machines combines VR teleoperation and large-scale simulation, with an option to train Moby in a customer’s facility twin when one is provided. The case study says Moby shipped to Asia within 18 months and two Moby3 units were brought to an unnamed semiconductor facility.
NVIDIA’s case study also says the integrated development workflow reduced an expected four-year programme with at least 50 people to 18 months with a team of about 15. NVIDIA presents the comparison as an initial estimate and publishes no method for calculating the four-year baseline.
The case study describes Moby’s current stack but does not identify the exact versions used during the original 18-month build. Noble announced its first industrial robot delivery on 3 March 2026, saying it had reached an unnamed Fortune Global 500 customer within 18 months. NVIDIA publicly released GR00T 1.7 on 7 July 2026. Public material does not establish whether version 1.7 contributed to the original milestone.
By 18 March 2025, Noble’s predecessor said its pipeline used Isaac Sim, Isaac Lab and Isaac GR00T synthetic-data tools. The account predates the March 2026 deployment announcement, although Noble has not disclosed the delivery date or exact software versions used. The company, then called Under Control Robotics, also said four engineers had built and deployed an earlier Moby system in under eight months.
VR Captures Human Demonstrations
Noble uses VR-based teleoperation to collect human demonstrations for robot training. An operator shows Moby how to perform a physical task, and those demonstrations are used to refine the GR00T 1.7 vision-language-action model.
NVIDIA says Noble uses VR-based teleoperation to collect real-world demonstrations for GR00T 1.7, alongside Isaac Lab for whole-body-control training and sim-to-real deployment. The case study doesn’t identify the headset, controllers or teleoperation application, quantify the VR-collected data, or say which semiconductor tasks used it.
Noble’s May account of its learning stack says two GTC tasks used one model trained on roughly ten hours of demonstration data. It separately says teleoperated expert demonstrations seed tasks and reports roughly tenfold scene and background augmentation. Across two supervised one-hour deployment cycles, operators collected about 20 minutes of targeted correction data. Noble does not say whether the original ten hours were collected through VR.
VR functions as a task-demonstration interface in Noble’s pipeline. The XR Beat’s earlier report on Tesla’s purchase of a Virtuix treadmill could only establish that the Optimus team had bought one dedicated VR system. Noble and NVIDIA now describe a working route from human demonstrations into robot-policy training, although the hardware and data volumes still need clarification.
Customer Twins Can Prepare Site-Specific Work
Noble imports Moby’s mechanical design into Isaac Sim and builds training environments from simulation assets. When a customer supplies a digital twin of its facility, engineers can place Moby’s digital counterpart inside the model before running the same work on physical hardware.
The Newton physics engine simulates contact-heavy actions such as lifting, grasping and moving objects. Isaac Lab trains whole-body control, while NVIDIA GPUs run thousands of digital Moby instances in parallel. Noble also works with Schaeffler on high-fidelity actuator twins that model friction, manufacturing variation and wear.
Industrial buyers would need to supply an accurate facility model or fund its creation. They also need an agreed refresh process because racks, routes, machinery, lighting and storage layouts change. The XR Beat’s reporting on digital-twin freshness covered the same operating constraint: simulation quality depends on how closely the model follows the live site.
Noble says its robots can learn skills in hours through language instructions, demonstrations and gestures. Public material does not disclose how long the semiconductor adaptation took, how much engineering work the customer supplied or whether its facility twin was used.
Onboard Compute Reduces Cloud Dependence
The current Moby uses a Jetson Thor development kit to process camera and sensor data and run foundation-model inference onboard. NVIDIA says this lets Moby run model inference in real time without external communication.
ADLINK and Noble are also developing a dedicated industrial computer built around Jetson Thor, with customised sensor connections, shock and vibration resistance and support for wider operating temperatures. ADLINK’s March announcement describes the dedicated system as the next step beyond its current DLAP support.
Onboard inference still requires clear answers on telemetry, software updates, cyber security, remote intervention, data retention and failure recovery. Noble’s website offers platform access for customers developing their own uses and limited robot-as-a-service pilots operated by Noble. Platform access places more integration and support work on the buyer, while Noble-operated pilots keep more responsibility with the supplier.
Two Robots Don’t Establish Production Performance
NVIDIA says Noble shipped Moby to Asia within 18 months of founding and brought two Moby3 units to a semiconductor facility for material-handling workflows. Noble’s March launch announcement separately referred to a deployment with an unnamed Fortune Global 500 industrial customer. Public material does not confirm whether both statements describe the same site.
Solomon, a Taiwan-based systems integrator, is helping adapt Moby for semiconductor material handling. Schaeffler supplies actuator expertise and ADLINK supplies edge-computing hardware. NVIDIA’s case study lists all three companies in its “Customer” field, while the body describes them as collaborators. None is identified as the operator of the semiconductor facility.
No published site-level figures for the disclosed semiconductor work cover uptime, throughput, intervention rate, autonomous task completion, actual shift duration, safety performance or return on investment. The commercial status is also unclear: the units could be in paid production, a pilot or an evaluation.
Industrial buyers have enough detail to inspect the development method, but they still lack the operating figures needed to compare Moby with fixed automation, autonomous mobile robots or other humanoids. Noble and its customer would need to publish the tasks, site-preparation time, human support requirement and production results before the semiconductor work can support a wider deployment claim.
