CONFIRMED
NVIDIA announced a major expansion of its Isaac open robotics development platform at GTC Berlin on Oct. 21, 2026, adding a reference humanoid, datasets, a runtime and updates to its robot learning framework, according to the company’s press release. Two days later, a NVIDIA technical blog showed AI agents setting up a simulated obstacle course for the small Microduck robot. Nine robot companies are named as partners. None of the material shows any of them running Isaac in production.
What we know
The hardware centerpiece is Agile ONE SN from Agile Robots. NVIDIA’s release describes it as a full-stack reference humanoid preconfigured with NVIDIA Isaac GR00T and powered by the Jetson Thor edge AI platform. NVIDIA says it unifies simulation, training and deployment tools on one platform and cuts setup time from weeks to days. That is NVIDIA’s own figure. We found no independent test of it.
The partner list needs a careful read. NVIDIA’s release says its open source robot learning reference workflows running on DGX Station “will be supported by solutions from” 1X, Agile Robots, Humanoid, Maven Robotics, Mentee, MicroAGI, Rhoda, Sereact and Skild AI. That is a statement about future support for development workflows. It does not say any of these companies ship robots running on the stack today.
Jensen Huang said in the announcement that the “ChatGPT moment for robotics is fast approaching.” That is a prediction, not a measurement.
The technical blog, “Using AI Agents to Teach Robots Like Microduck to Walk in Simulation,” is dated Oct. 23, 2026. In it, NVIDIA RTX Spark serves as a local hub where AI agents such as Hermes and Nemotron 3.5 Bolt turn a prompt into a step-by-step project plan. Astra builds the 3D assets and exports them as USD. The course is assembled in Isaac Sim, with Newton handling the physics of Microduck’s movement and balance.
Isaac Lab then trained a perception and locomotion policy in 4,096 parallel environments on a single RTX PRO 6000 Blackwell Workstation Edition GPU, using head-camera depth plus robot-state observations. The policy was validated in Isaac Sim on RTX Spark, with sim-to-real preparation using Better Actuator Models and domain randomization. Every obstacle-course result in the post is a simulation result.
What we don’t know yet
Some of the most interesting pieces are not yet confirmed. We could only partly retrieve NVIDIA’s full press release, so a few details rest on secondary trade coverage. The AI Insider and Blockchain.News report that the expansion makes Isaac Lab 3.0 generally available and introduces Isaac OS, an open-source runtime linking AI models to sensors, controls and robot hardware. They also report a SimReady Central Library of more than 2,000 assets, a Human 1M Dataset, and new DGX Station workflows for GR00T, dexterous manipulation and Cosmos world-model training. Treat these as reported until checked against the full release.
On timing, Blockchain.News reports that Agile ONE SN is scheduled for early 2027 and that Agile Robots plans to integrate NVIDIA Halos safety capabilities. Engineering.com reports Jetson Thor specs of 2,070 FP4 teraflops and 128 GB of unified memory. No source we found gives a price for Agile ONE SN.
We also don’t know how Microduck performs on the obstacle course in the real world. NVIDIA’s blog stops at preparing for sim-to-real transfer, and Blockchain.News notes that physical deployment is outside the post’s scope. CTOL Digital reports that physical Microduck demos show walking only, while the obstacle-course runs remain in simulation. No source describes any demo as teleoperated or sped up.
What it means
For humanoid developers, the concrete new items are a reference robot that arrives with GR00T and Jetson Thor already integrated, and a documented agent-assisted pipeline from prompt to trained policy in simulation. If the reported runtime and Isaac Lab 3.0 hold up against the full release, the stack moves closer to an end-to-end toolchain.
The open question is the last step. Intelligent HQ’s coverage points out that sim-to-real transfer remains a hard problem, because differences in sensors, friction and object properties still require validation and testing. Training 4,096 simulated robots on one GPU is impressive throughput. Getting one physical robot through the same course is a separate claim, and nobody has made it yet.
As for production adoption, the honest answer from the available sources is none confirmed. Nine companies have signed on to support NVIDIA’s workflows. That is a meaningful ecosystem signal. Robots on factory floors running this stack would be evidence of a different kind, and that evidence has not been shown.
Follow every development in one place on our humanoid robot tracker.
Keep reading
Robotics NewsOptimus Academy needs 10,000 to 20,000 robots, Musk says, but self-play is unprovenMusk says Tesla may need 10,000 to 20,000 Optimus robots to practice tasks. It is a plan, not…Oct 23, 2026 · 4 min read
Robotics NewsUnitree H2-X kung fu video appears real but staged; live-demo fall report is unconfirmedUnitree's H2-X video appears real but staged. A live-demo fall is unconfirmed, the car lift is questioned, and…Oct 23, 2026 · 4 min read
Robotics NewsUnitree’s 2028 humanoid target is conditional, and self-improving robots are unprovenUnitree's founder floats 80% of tasks by voice as soon as 2028. His self-improving robot loop is still…Oct 23, 2026 · 4 min read
