NVIDIA Isaac at GTC Berlin: Agile ONE SN, the ready-to-use reference humanoid

NVIDIA Isaac at GTC Berlin: Agile ONE SN, the ready-to-use reference humanoid

🔎 A reference humanoid, and why now

On October 21, 2026, at GTC Berlin, NVIDIA turned a problem that robotics has been dragging around for years into a simple checkbox. The Isaac platform is gaining "reference humanoid frameworks," simulation pipelines, and open datasets. And the first official body of this initiative has a name: Agile ONE SN, developed with Agile Robots.

The problem being addressed is familiar to anyone who has ever launched a robotics project. Between hardware selection, simulation setup, model adaptation, and deployment, weeks — sometimes months — go by before the first useful experiment. According to Unite.AI, the stated goal is precisely to cut this setup time from weeks to days.

Jensen Huang summed up the ambition in a single sentence: "the ChatGPT moment for robotics is approaching." Translation: action models are maturing, massive data is arriving, and what's missing is a standard body everyone can agree on. That's exactly what NVIDIA just delivered.

The Essentials

• Agile ONE SN: reference humanoid, 1.40 m tall, 40 kg (43 kg per the manufacturer's spec sheet), 31 degrees of freedom in the body, 3 kg payload per arm, WUJI Hand 2 five-finger hands with 20 DoF and 1 kHz force control (NVIDIA, October 2026).
• Onboard compute: Jetson Thor with 2,070 TFLOPS in FP4 and 128 GB of unified memory.
• Platform: Isaac OS runtime, Isaac Lab 3.0, SimReady library with over 2,000 simulation assets, Human 1M Dataset on Hugging Face.
• Model: GR00T 1.7, an open source VLA preconfigured on the SN, with URDF and USD assets for Isaac Sim and Isaac Lab.
• Availability: early 2027, with interest already expressed by Stanford and ETH Zurich.
• Our take: the product isn't the robot, it's the pipeline.

Recommended Tools

Agile ONE SN: the specs of a humanoid designed as a standard

The Agile ONE SN is a bipedal humanoid measuring 1.40 m and weighing 40 kg, with 31 degrees of freedom in the body, designed by Agile Robots to serve as the official reference hardware for the NVIDIA Isaac platform.

The idea of a "reference body" is nothing trivial. Today, every lab training a robotics model does so on its own hardware: learned policies don't transfer, benchmarks can't be compared, and every team pays the same integration toll. By imposing a common, documented, and simulable body, NVIDIA hopes to create a de facto standard.

The choice of dimensions is not neutral either. At 1.40 m, the SN operates in environments designed for humans: door handles, stairs, workbenches, standard tools. It's a deliberate compromise between the maneuverability of an office robot and the strength of an industrial machine.

On the actuation side, the SN doesn't do things by halves. Its WUJI Hand 2 hands each have 20 degrees of freedom across 5 fingers, with force control at 1 kHz — the frequency needed to grasp a fragile object without crushing or dropping it. Joint torque reaches up to 140 Nm according to the manufacturer's spec sheet (Agile Robots, October 2026).

Here is the consolidated technical specification sheet:

40 or 43 kg? Two spec sheets, two truths

Yes, both figures are circulating, and it's instructive. NVIDIA's press release announces 40 kg and 31 degrees of freedom; the manufacturer's spec sheet from Agile Robots lists 43 kg and 71 degrees of freedom in total, hands included.

The difference in degrees of freedom is easily explained: 31 in the body, plus 2 × 20 for the WUJI Hand 2 hands, gives exactly 71. Both figures are true — they just don't count the same thing.

As for the weight, the 3 kg gap most likely comes down to the measured configuration — batteries, options, cladding. The lesson to remember: when facing a press release, the manufacturer's spec sheet is the authoritative source. We're giving you both — it's the bare minimum.

Another point of honesty: 3 kg of payload per arm is modest. We're far from an arm carrying a 10 kg crate or heavy tooling. The SN is clearly positioned as a research and development platform, not a factory worker. Agile Robots also points out that skills acquired on the SN transfer to the Agile ONE, its industrial version — the SN is the training ground, the ONE is the production machine.

Jetson Thor and Isaac OS: embedded compute, the real heart of the announcement

The real product of this announcement isn't the robot: it's Jetson Thor and the Isaac OS runtime, which finally make it possible to run a full VLA model on the robot itself.

The numbers set the tone: 2,070 TFLOPS in FP4 and 128 GB of unified memory on an embedded module (NVIDIA, October 2026). Why is that crucial? Because vision-language-action models are resource-hungry: their weights and activations exceed what most embedded platforms can host. With 128 GB of unified memory, the model lives on the robot, with no CPU-GPU copies and no network latency.

This is the logical next step in NVIDIA's hardware strategy. When the company unveiled Vera Rubin and the ARM N1X chips at GTC Taipei, it showed it was thinking about AI from cloud to edge — see our analysis of NVIDIA GTC Taipei: Vera Rubin, N1X ARM and the era of AI agents. Jetson Thor is the robotics take on that vision: the same ecosystem, from the datacenter all the way down to the joints.

The Isaac OS runtime completes the picture by unifying the software layer across simulation, training, and deployment (GlobeNewswire, October 2026). And NVIDIA plans to integrate its Halos safety stack — a strong signal. When a 40 kg robot moves near humans, safety isn't a marketing argument, it's an architectural building block.

My take: this is where NVIDIA's real game is being played. A robot without a software ecosystem is just an expensive servomotor. With Thor, Isaac OS, and Halos, NVIDIA is selling a complete environment — exactly the recipe that worked for the gaming PC and the datacenter.

Isaac Lab 3.0, SimReady, and Human 1M: The Real Revolution Is in the Data

The least publicized part of the announcement — the data and pipelines — is paradoxically the most important for the future of general robotics.

Three building blocks arrived together. Isaac Lab 3.0, the new version of the open source training framework, covers reinforcement learning and imitation learning with sim-to-real pipelines. The SimReady Central Library makes over 2,000 ready-to-use simulation assets available — physically consistent objects and environments, where teams used to spend weeks modeling them. Finally, the Human 1M Dataset, published on Hugging Face, provides a massive corpus of human movements (GlobeNewswire, October 2026).

Context matters. NVIDIA and Hugging Face have already joined forces in LeRobot, with the open source release of Isaac GR00T 1.7, Isaac Teleop, and Cosmos 3 — a partnership we covered in NVIDIA et Hugging Face unissent leurs forces dans LeRobot. The Human 1M Dataset fits into this momentum: training data is becoming a common good rather than a private asset.

Why is this decisive? Because robotic data is the sector's budgetary abyss. Collecting real-world demonstrations requires hours of teleoperation, robots, and technicians. A million human movements combined with synthetic data changes the economics of training.

Unite.AI summarizes the effect: unifying simulation, training, and deployment to bring setup time from weeks down to days (October 2026). Concretely, a PhD student can clone a SimReady environment, train a policy in Isaac Lab 3.0, and test it on a reference body — without ever having to negotiate with a procurement department.

This is exactly what made NLP take off: public corpora before the models. Robotics has just received its data infrastructure. It's less sexy than a humanoid doing push-ups, but it's what will decide who wins the race.

GR00T 1.7 and open workflows: when the model meets the body

GR00T 1.7 is an open source vision-language-action model, preconfigured on the Agile ONE SN, and it's what turns this reference robot into a learning platform that's immediately usable.

The preconfiguration changes everything. Agile Robots provides URDF and USD assets for Isaac Sim and Isaac Lab, and the GR00T 1.7 VLA comes preconfigured on the SN (Agile Robots, October 2026). No more months of engineering to match a generic model to a specific body: you download, you train, you deploy.

The software architecture is layered, and that's by design. At the bottom, the GR00T 1.7 VLA handles the perception-action loop: see, understand, act. On top of it, the announced open source workflows combine ChatGPT Astra, Isaac and Cosmos on a DGX Station (GlobeNewswire, October 2026) — general-purpose large models like OpenAI's GPT-5.5 or Google's Gemini 3.1 Pro handle reasoning and task orchestration, while the VLA executes the movements.

Cosmos 3 plays its role in this chain: generating realistic synthetic data to train the policies without multiplying teleoperation hours. We analyzed this building block in NVIDIA Cosmos 3 and Isaac GR00T: the ChatGPT moment of robotics — and Jensen Huang's formula, "the ChatGPT moment of robotics is approaching," takes on its full meaning here.

My blunt take: preconfiguring an open source VLA on a reference body is the most strategic move of the announcement. It creates a controlled coupling between Agile Robots' hardware and NVIDIA's software — open enough for researchers, integrated enough for industry. That's the very definition of a de facto standard.

Agile ONE SN vs RoboParty RP1: industrial reference or fully open source?

The Agile ONE SN and the RoboParty RP1 answer the same question — "how do you start a humanoid project without reinventing everything?" — with two opposing philosophies, and both are right.

The RP1 fired the first shot by releasing its hardware and software — we covered it in RoboParty RP1: the 100% open source humanoid that opens up hardware and software. The SN answers with a different weapon: a validated body, a preconfigured model, and an industrial ecosystem behind it.

The analogy that works: the SN is Android — an open but orchestrated standard, with the guarantees of a manufacturer. The RP1 is the community distribution — total freedom, support is up to you. The two will coexist, and that's excellent news for robotics.

My advice depends on your profile. If you're a lab that wants to publish comparable results, or a company that wants an accountable point of contact, the SN is for you. If you want to understand every line, tinker with the hardware, and depend on no vendor, the RP1 remains unbeatable.

Blind Spots: What the Press Release Doesn't Say

Three gray areas deserve to be put on the table before the order book is released.

Dependence on the NVIDIA stack, first. Thor, CUDA, Isaac OS, Halos: every brick reinforces the lock-in. The open-source components — Isaac Lab, GR00T 1.7, the workflows — soften the pill, but the optimization will remain calibrated for NVIDIA hardware. That is the price of the standard.

Economic opacity, next. Neither the price of the SN, nor that of Jetson Thor, nor the licensing model of Isaac OS has been communicated (October 2026). A lab must budget for the robot, the workstation, and potentially the services. Without a price list, it is impossible to seriously compare it with a fully integrated open-source alternative.

Competition, finally. The transfer of skills from the SN to industrial Agile ONE is also a retention strategy: you train on the reference body, you buy the production machine. Teams that want to maintain their independence should closely follow the open-source ecosystem, where LeRobot and community projects are progressing quickly.

Early 2027: Who Should Wait, Who Can Start Today?

The robot itself won't arrive until early 2027, but almost the entire stack is already usable — and waiting for the hardware to build up skills is a strategic mistake.

NVIDIA already cites interest from Stanford and ETH Zurich (October 2026). Two institutions that generally don't commit to a press release without having vetted the goods. This signal matters: the SN isn't vaporware, it's a program with identified users.

Here's what you can do today, without spending a dime:

1. Install Isaac Sim and Isaac Lab 3.0, both open source, to get familiar with sim-to-real pipelines.
2. Download the Human 1M Dataset on Hugging Face and explore the human motion data.
3. Study GR00T 1.7 and its URDF/USD assets provided by Agile Robots to understand the architecture of a preconfigured VLA.
4. Follow the open source workflows combining ChatGPT Astra, Isaac, and Cosmos, documented for DGX Station.

For teams with a budget, the question is different: should you pre-order an SN? My take: wait for pricing and support terms, but start planning now for a budget covering development on Jetson Thor. The robot will be there in early 2027; your team's skills, on the other hand, are built now.

❌ Common Mistakes

Mistake 1: Confusing "reference humanoid" with "deliverable product"

The announcement is about a reference framework, not a product in stock. Availability is announced for early 2027 (NVIDIA, October 2026). The solution: start developing on simulation now and consider the hardware as the final step, not the first.

Mistake 2: Taking specs from the press release without checking the manufacturer's sheet

40 kg or 43 kg? 31 or 71 degrees of freedom? Both numbers are correct, but they don't measure the same thing—the configuration and counting scope differ. The solution: systematically cross-reference the NVIDIA press release with the Agile Robots sheet, and cite both.

Mistake 3: Believing that simulation eliminates real-world validation

Isaac Lab 3.0 and the 2,000+ SimReady assets reduce the sim-to-real gap; they don't eliminate it. The 1 kHz force control of the WUJI Hand 2 exists precisely because real contact remains unpredictable. The solution: budget for physical validation hours, and integrate the Halos stack for any deployment close to humans.

Mistake 4: Waiting for the robot to train on the stack

This is the most costly mistake. Isaac Sim, Isaac Lab 3.0, GR00T 1.7, the Human 1M Dataset: everything is open source and available since October 2026. The solution: upskill on the software side now, so that you arrive in early 2027 with a trained pipeline rather than with curiosity.

❓ Frequently Asked Questions

When will the Agile ONE SN be available?

NVIDIA announced availability in early 2027, with interest already expressed by Stanford and ETH Zurich (October 2026). No precise date or pre-order policy has been communicated. Monitor the NVIDIA and Agile Robots pages: the cited laboratories will likely receive the first evaluation units.

How much does the Agile ONE SN cost?

No price has been communicated to date (October 2026). NVIDIA has not published any figures, and we will not speculate on a market where price differences are measured in tens of thousands of euros. Check nvidia.com and agile-robots.com at the time of availability, early 2027.

Why does NVIDIA announce 31 degrees of freedom and Agile Robots 71?

Both figures are accurate: NVIDIA counts the 31 degrees of freedom of the body, while Agile Robots adds the 2 × 20 degrees of the WUJI Hand 2, totaling 71. This is a matter of counting scope, not a contradiction. For engineering specs, refer to the manufacturer's datasheet.

Can one develop for Isaac without buying a robot?

Yes, and that is the whole point of the announcement. Isaac Sim and Isaac Lab 3.0 are open source, the SimReady Central Library offers over 2,000 assets, the Human 1M Dataset is available on Hugging Face, and GR00T 1.7 is pre-configured on the SN. Only the physical deployment will wait until early 2027.

What is the purpose of the 128 GB unified memory of Jetson Thor?

Vision-language-action models are large, and unified memory avoids copies between the CPU and GPU that slow down latency. With 2,070 TFLOPS in FP4, Jetson Thor runs the VLA directly on the robot. This is the condition for an embedded perception-action loop, without cloud dependency.

Is the NVIDIA Isaac stack really open source?

Partially, and you must read the licenses component by component. Isaac Lab 3.0, GR00T 1.7, the workflows, and the Human 1M Dataset are open; the Isaac OS runtime has been announced as a platform addition; the Halos security stack is a product integration. The openness is real but orchestrated.

✅ Conclusion

NVIDIA isn't selling just another robot: it's imposing a standard body, a unified runtime, open data, and a pre-configured VLA — the complete pipeline that general-purpose robotics was missing. The ChatGPT moment for robotics isn't here yet, but for the first time, we can see the mechanics that will get us there. Start with Isaac Sim and the Human 1M Dataset: early 2027 will come around fast.