NVIDIA releases Isaac Lab 3.0 and sets out an early-2027 humanoid reference platform

Isaac's expansion connects robot learning, runtime software and simulation assets, while Agile ONE SN is scheduled for research availability next year.

NVIDIA expanded its Isaac robotics platform on 21 October with general availability of Isaac Lab 3.0, a new Isaac OS runtime and a research humanoid announced by Agile Robots. The package spans training data, simulation, policy learning and deployment, reflecting the amount of integration required to move a robot capability from a demonstration to a repeatable system.

Agile ONE SN, the humanoid reference platform, is scheduled for early 2027. NVIDIA describes a 1.4-metre, 40-kilogram body with 31 degrees of freedom, five-finger hands, multimodal sensing and Jetson Thor onboard computing. Integration with the Halos safety stack is planned.

A common starting point for research

A reference robot gives researchers a shared set of mechanical and computing assumptions. The benefit is practical: work on manipulation, perception or coordinated movement can begin from a documented platform instead of a fresh combination of actuators, cameras and control software.

It can also make experiments easier to compare. A policy trained for one body still has to match that body's reach, balance, payload and sensing. Recording those details beside the learned behaviour helps explain whether a result transfers to another machine or depends on a particular physical configuration.

The distinction between the available software and the planned robot is clear in the announcement. Teams can begin examining Isaac Lab 3.0 now, while the reference hardware has a separate delivery schedule. Interest from research institutions is described as interest rather than completed deployments.

Learning tools meet the runtime

Isaac Lab 3.0 supports robot-policy training across multi-GPU workflows and physics engines including Newton and PhysX. NVIDIA describes Isaac OS as an open runtime framework built on Holoscan to connect AI, sensors, control and hardware with real-time behaviour.

That connection addresses a difficult handoff. A training environment advances in simulation steps; a physical machine receives sensor data, schedules computation and moves motors against real time. Keeping the interfaces explicit lets a developer identify which part of a behaviour belongs to the learned policy and which part belongs to the controller or runtime.

The release includes reference workflows for refining GR00T models with reinforcement learning, training dexterous manipulation and adapting Cosmos world models. A SimReady library contains more than 2,000 robotics and factory assets, and the Human 1M dataset combines human activity with synchronised views and hand tracking. Access to the full dataset is described as available to Cosmos Coalition partners.

Venture investor Amulya Vishwanath calls physical AI a “LEGO era”: assembling useful components before a broader breakthrough.

More than a single robot launch

The announcement identifies industrial partners using different parts of the stack for virtual commissioning, goods handling, perception and mobile-robot development. That variety highlights the platform's modular character: a manufacturer may need a simulation pipeline without adopting the humanoid, while a research group may want the robot body and training framework together.

For developers, the useful unit of progress is a complete behaviour with known inputs, a repeatable training setup and an operating interface on the target machine. The expanded Isaac offering supplies more of those connecting components.

The next stage will be visible in how teams use the available software and how the scheduled reference hardware reaches researchers. For now, the concrete change is a larger shared development environment, with immediate software releases and a separately dated hardware roadmap.

Sources

- [NVIDIA Advances Isaac Platform for General-Purpose Robotics](https://nvidianews.nvidia.com/news/nvidia-advances-isaac-platform-for-general-purpose-robotics-with-new-reference-humanoid-frameworks-simulation-pipelines-and-data) — NVIDIA