Technology
Agile Robots Unveils ONE SN Research Humanoid for Early 2027
Agile Robots introduced the 1.4-meter Agile ONE SN research humanoid on October 21, combining dexterous hands, onboard NVIDIA computing and a preconfigured robot-learning workflow. Availability is scheduled for early 2027; the partners' launch releases disagree on mass.
Agile Robots has introduced Agile ONE SN, a 1.4-meter research humanoid that packages its robot hardware with NVIDIA's computing and robot-learning tools. Announced at GTC Berlin on October 21, the system is scheduled for availability in early 2027.
The proposition is practical: give a laboratory a robot, simulation assets and a learning workflow that already fit together. The announcement adds a defined developer product to Agile's humanoid strategy, although pricing, customer deliveries and sustained autonomous task results are still undisclosed.
Key specifications announced
1.4 m
Height
71 DOF
Body and two hands combined
3 kg
Payload per arm, up to
2027
Availability scheduled early in year
A development platform with a defined starting point
A humanoid research project can spend substantial effort connecting cameras, motors, simulators and learned policies before testing a useful behavior. Those interfaces matter: a controller expects particular sensor timing, joint definitions and action formats. Changing any one can affect the rest of the experiment.
Agile ONE SN targets that integration work. The company's launch announcement describes preconfigured Isaac GR00T software, including the GR00T 1.7 vision-language-action model and whole-body control. A vision-language-action model connects visual observations and instructions with robot actions; whole-body control coordinates the robot's movements.
The product page sets out a sequence of collecting demonstrations, training, evaluating in simulation and testing on the physical robot. URDF and USD assets describe the robot for simulation tools. Having those descriptions supplied can make experiments easier to reproduce, provided the models accurately represent the delivered hardware.
That is a useful purchasing proposition for a lab: a common starting configuration can let researchers spend more time studying behavior. It also creates a testable promise. Teams can measure installation hours, calibration effort and the time required to reproduce a supplied task before accepting claims of easier development.
Under the hood: the body, hands and a weight discrepancy
Agile specifies 31 body degrees of freedom and two Wuji hands with 20 each, producing the advertised combined count of 71. Degrees of freedom describe independently controllable movement; the number alone says little about how reliably a robot can complete a task.
The two partners disagree on mass. Agile's announcement gives 43 kilograms, while NVIDIA's release gives 40 kilograms. The reviewed announcements do not reconcile those figures. Buyers need a configuration-specific mass, including a clear statement about hands and other installed equipment.
| Specification | Agile Robots announcement | NVIDIA announcement |
|---|---|---|
| Height | 1.4 meters | 1.40 meters |
| Mass | 43 kilograms | 40 kilograms |
| Body movement | 31 DOF | 31 DOF |
| Hands | Two tactile five-finger Wuji hands, 20 DOF each | Two WUJI Hand 2 units, 20 DOF each |
| Arm payload | Up to 3 kilograms per arm | 3 kilograms per arm |
| Availability | Beginning of 2027, scheduled | Early 2027, scheduled |
The sensor layout includes head and wrist cameras and lidar, which measures distance using light. The product page identifies Jetson Thor for onboard inference and sensor processing, complemented by an Intel i3 N305 computing module. These components establish an architecture; they do not establish the latency or reliability of a complete learned behavior.
Likewise, the arm payload needs operating conditions. Reach, acceleration, grasp stability and the object's shape all affect a real lifting task. A laboratory evaluating the stated limit should ask which motions and postures it covers.
NVIDIA identifies the hands as WUJI Hand 2. Wuji's own specification page lists 20 active axes, a 1,000-hertz control rate and backdrivable rotary actuation, while warning that its Beta1 specifications continue to change. Component control frequency should not be read as the decision rate of the robot's vision-language-action model.
Agile Robots: connecting a research product to an industrial business
Agile ONE SN is the compact developer platform in a portfolio that also includes the full-size Agile ONE. Agile says skills and data developed on SN can later transfer to that industrial robot. Such transfer needs demonstration: differences in reach, mass, hand geometry and balance can change whether a learned action works.
The group includes Franka Robotics, idealworks, BÄR Automation and Krause Automation. Its industrial background offers experience with applications and integration. Historical financing provides another part of the context: the company's September 2021 Series C announcement reported $220 million, led by SoftBank Vision Fund 2. That is historical funding, not a new round attached to ONE SN.
The launch therefore advances the question raised by Biped's earlier reporting on Agile's industrial strategy: what concrete humanoid product can grow from that base? SN supplies a research configuration and availability target. The next evidence must come from delivered systems, repeatable experiments and support in users' laboratories.
Agile and NVIDIA cite interest from Stanford University and ETH Zurich. Interest is an early demand signal; the announcements do not identify delivered units or completed research programs at either institution. Institutional names should be evaluated alongside what was actually supplied and tested.
What this means for laboratories
A common hardware and software platform could make results easier to compare. Two teams using the same robot configuration can investigate differences in data, training and evaluation with fewer hardware variables. Shared simulation assets can also make failures easier to reproduce before anyone risks damaging a physical machine.
Procurement still needs details beyond the specification sheet. Relevant questions include the supplied software versions, access to calibration tools, permitted modifications, model and dataset licenses, replacement parts and the support available when a demonstration fails. An open and modular description does not by itself specify the license or warranty for every component.
Researchers also need to record interventions. A long task completed after several human resets answers a different question from the same task completed autonomously. Reporting the number of trials, resets, falls and changed objects would make demonstrations more informative than a selected successful video.
For a manipulation experiment, the useful unit of comparison is the entire attempt: perceive the object, approach it, grasp it, move it and recover when something goes wrong. A robot might complete each individual motion in isolation yet struggle when their errors accumulate. Reporting intermediate progress alongside complete-task success would reveal where the system needs improvement.
The physical setting matters too. Wrist-camera views can change when an arm blocks the scene, and object placement can change the reachable grasp. A repeatable evaluation should document those conditions rather than silently simplify them between runs. Supplied simulation assets make this kind of comparison easier to organize, but the physical tests still need to show how closely the model matches reality.
Deployment reality check
ONE SN is an announced research platform presented for GTC, with early-2027 availability planned. The reviewed launch materials do not disclose customer shipments, operating fleets, sustained task-success rates or human-intervention rates.
The emergency stop and planned Halos integration are relevant design features. Agile describes Halos integration as exploratory, and NVIDIA describes it as planned. Neither statement supplies a completed system safety assessment or authorization for unrestricted operation around people.
The milestones before early 2027
The first milestone is a clear orderable configuration: price, delivered mass, included hands, software package and delivery commitments. The 40-versus-43-kilogram discrepancy is small enough to resolve explicitly and significant enough for handling, balance and simulation assumptions.
The next milestone is reproducibility. A documented task suite should specify starting conditions, success criteria, trial counts and human assistance. Cross-platform transfer to full-size Agile ONE would be particularly useful evidence if the company reports both the adaptation required and the remaining failures.
A final commercial question is support. Research hardware produces value when teams can keep experimenting through worn parts, software updates and calibration changes. Service response and the cost of maintaining a stable configuration will influence whether SN becomes a lasting laboratory tool.
Frequently asked questions
When will Agile ONE SN be available?
Agile Robots and NVIDIA schedule availability for early 2027. The reviewed launch announcements do not disclose a price or confirm customer deliveries.
How many degrees of freedom does it have?
Agile lists 31 body degrees of freedom and two hands with 20 each, totaling 71. That movement count does not measure task success or autonomous reliability.
Does Agile ONE SN weigh 40 or 43 kilograms?
Agile's October 21 announcement says 43 kilograms; NVIDIA's says 40 kilograms. The reviewed announcements do not explain the difference, so a delivered configuration's mass still needs confirmation.
Is Halos safety integration complete?
The partners describe Halos integration as planned or being explored. The reviewed launch materials do not establish completed integration or a certified safety case for the whole robot.
A useful platform will make its experiments reproducible
ONE SN gives Agile's humanoid program a more concrete entry point for researchers: specified hardware, a development workflow and a delivery window. Its value will depend on how well those parts work together in laboratories beyond the launch demonstration.
The decisive evidence will be researchers reproducing useful behaviors on delivered systems, with the setup work, failures and human assistance reported alongside the successes.
Image context: the hero is an archival Franka Emika arm photographed at CeBIT on March 23, 2017, illustrating the industrial robotics background of Agile's wider portfolio. It does not show ONE SN or the October 2026 launch. Photograph by Ims, via Wikimedia Commons, CC BY-SA 4.0. Resized rendition and responsive display crop; image adaptations remain under the same license.
