Agile ONE SN gives robotics researchers a common starting point

Agile Robots and NVIDIA have introduced a humanoid intended to give research teams a consistent machine on which to develop robot skills. Announced on 21 October, Agile ONE SN comes preconfigured with Isaac GR00T and Jetson Thor computing, with availability scheduled for early 2027.[1]

The configuration combines a 1.4-metre body, two five-finger hands with tactile sensing, and cameras at the head and wrists. Alongside it, NVIDIA announced Isaac Lab 3.0 for learning in simulation and Isaac OS, a runtime connecting AI software, sensors, controls and hardware.[1]

The intended audience is visible in Agile chief executive Zhaopeng Chen’s interview with Reuters: universities and corporate research laboratories. Chen said production of the ONE S and ONE SN models would begin by the end of 2026, with 1,000–2,000 machines expected in 2027 across the two models. Those are production plans; the figure does not describe robots already delivered to workplaces.[2]

A robot experiment has several moving parts

The value of a common configuration becomes clearer when following a skill from training to execution. A policy is the software that turns observations into actions. Its inputs might include joint positions or a camera image; its outputs must eventually correspond to actual movement. The same numerical output can have a different meaning if the receiving software expects a different joint order, scale or update interval.[4]

Isaac Lab’s deployment documentation makes that handover explicit. For supported manager-based training environments, it can export descriptions of a policy’s inputs and outputs, including joint names, their order and timing parameters. Another tool can use that description to reconstruct the configuration. It is a modest but consequential form of standardisation: the learned behaviour travels with information about how to interpret it.[4]

For a research group, reducing that translation work could make comparisons more meaningful. If two policies are tried on the same configured platform, a change in behaviour is easier to investigate than when the experiment also changes cameras, hands, control settings and software connections. That is an experimental advantage to pursue; the announcement does not itself measure how much time a particular laboratory will save.[5]

The virtual object needs more than a convincing appearance

NVIDIA’s new SimReady Central library addresses the environment around the robot. Its more than 2,000 downloadable assets cover robotics and industrial equipment. The important distinction is the information attached to each object: properties such as mass, collisions, friction and joints allow a simulator to calculate interactions. Published profiles describe which capabilities an asset supports, helping developers select material suitable for their particular experiment.[3]

A realistic-looking box is enough for an image, but a grasping experiment also needs to know how that box responds when a hand touches it. A shared library can remove repeated scene-building work while making those assumptions easier to inspect. It still matters whether the selected asset includes the properties required for the task; visual resemblance alone cannot answer that question.[3]

The documentation also puts useful boundaries around repeatability. Isaac Lab explains that rigid-body simulations can reproduce results with the same hardware and relevant software versions, while different hardware can introduce numerical differences. Its PhysX discussion does not promise determinism for deformable objects such as cloth. Researchers therefore need to preserve the experimental configuration as well as the trained policy when reporting a result.[5]

Seen together, the humanoid, asset library and deployment descriptions target a practical bottleneck: getting different parts of a robotics experiment to agree about the world they share. A laboratory still has to decide what skill to teach and how to evaluate it. A more consistent starting point could let researchers devote more of that effort to the behaviour they are trying to understand.[4][3]

Report an error or suggest a correction