NVIDIA’s GTC Berlin keynote clip of Hugging Face’s Microduck robot invites a big conclusion: that AI agents took a US$399 biped from simulation to the real world. NVIDIA’s own engineering write-up describes a narrower job, planning an obstacle course, building its props and assembling it in Isaac Sim, and it puts “detailed hardware deployment” outside its scope. That is a real step for agent tooling, but the recipe that makes a 25 cm robot walk on a real floor existed before the keynote.
What NVIDIA showed on stage
On October 21, near the end of his Berlin, Germany, keynote, NVIDIA chief executive Jensen Huang held up an RTX Spark, the PC platform NVIDIA released in laptops on October 16, with up to 128GB of unified memory and 1 petaflop of AI compute. He then played a video of an agent helping train Microduck. NVIDIA’s live recap says the agent set up a course, simulated the robot in Isaac Sim, sent training jobs to the cloud and brought the learned policy back to RTX Spark for evaluation “before deploying it to the physical robot.” Four Microducks then joined him on stage and sat, stood and napped on cue.
Huang told the ducks, “you learned your skills right here on RTX Spark and the digital twin.” NVIDIA’s RTX Spark account posted the clip with the caption “When @jensenhuang asks for a somersault, you deliver.” The post added that AI agents on RTX Spark “help bring” Microduck “from simulation to the real world.” TechCrunch reported on October 23 that Hugging Face chief executive Clem Delangue said in a tweet that preorders had passed 20,000 units.
What the agents did, in NVIDIA’s own words
Two days later, NVIDIA’s technical blog laid out the workflow. A prompt goes to Hermes Agent, an open-source agent from Nous Research, and to NVIDIA’s Nemotron 3.5 Bolt model, both running on RTX Spark. They turn the prompt into a step-by-step plan for an obstacle course. Hermes then delegates asset work to Astra, which builds the props in FreeCAD and Blender and composes the course in Isaac Sim, and Isaac Lab trains the robot’s policy. The trained policy returns to RTX Spark for evaluation in simulation.
The training step builds on existing work. The post says the experiment used “the Isaac Lab Microduck example environment as its starting point,” with a locomotion policy “pretrained in that environment” that then learned ramps and obstacles across 4,096 parallel environments. It also states that for each task a developer defines “what the robot observes, the actions it can take, and the rewards used to guide learning.” The post’s section on the physical robot carries a limit: “Detailed hardware deployment is outside the scope.”
NVIDIA’s separate Hugging Face tutorial, published October 21, covers the real robot. It has a developer train a walking policy with a typed command, export it as an ONNX file and copy it to the robot’s onboard computer. The word “agent” does not appear in it. The tutorial says Isaac Lab “already provides Microduck’s robot model, scene, observations, actions, rewards, and domain randomization.”
The implication is that the agents’ documented work sits at the front of the pipeline, in planning and asset creation, while the step that puts a policy on a real robot runs through a recipe someone had already built. The keynote recap and the technical post describe the same workflow but claim different scopes, and the technical post is the more careful of the two.
A somersault that predates the keynote
The somersault has a public history. Pollen Robotics, the Bordeaux, France, robotics team that joined Hugging Face in April 2025, lists a “Roulade” task in its training repository: a “forward roll over the head, land back on the feet.” The commit history shows that work landing between August 11 and August 18. Pollen’s official policy repository on Hugging Face has held a file named roulade.onnx since its first upload on August 31, four days after preorders opened.
Neither NVIDIA’s recap nor its technical post says which policy the stage ducks ran, so that file proves nothing about the keynote itself. It does show the behavior predates RTX Spark laptops, which launched on October 16. The task’s docstring carries the label “attempt 3, run 2” and records that the first run learned “a violent ballistic ‘breakdance’ whip,” which was “optimal under the run-1 rewards.” Among the fixes, rotation now counts only while the robot touches the ground.
Where sim-to-real still bites
Pollen’s README names the difficulty at this scale: “actuator fidelity is most of the sim2real gap.” Tiny servos drive a robot of under 800 grams, so the training simulation models each servo’s voltage control law, randomizes command delay and battery voltage sag, and, in its backlash variants, adds one degree of gear play either way on every joint. NVIDIA’s technical post leans on the same fix, a Better Actuator Model, then adds domain randomization because “sensors, ground contact, and control timing can differ on the real robot.”
The same repository shows what that gap costs in practice. Its AGENTS.md file, in the repository since launch week, opens by saying that every convention in it “exists because breaking it produced a policy that worked in the viewer and failed on hardware.” A centre-of-mass randomizer that accumulated across resets “degraded every long run for months.” A target standing height wrong by 5 millimetres “turned the goal into an impossible target for days.”
The README says the guide is “also aimed at AI coding agents working in this repo,” and agents already work there: 96 of the repository’s 1,255 commits carry a Claude co-author trailer. One reading is that agents become useful for sim-to-real once people have written down what breaks. NVIDIA’s technical post does not show an agent discovering a failure mode like those above, which by Pollen’s own account can cost days to months.
The heavy compute lives elsewhere
The duck’s own computer is small. Pollen’s press kit lists a Rockchip RK3566 processor, from China-based Rockchip, with 1GB of RAM, 32GB of storage and a removable battery good for around one hour of runtime. A 50 Hz loop executes the policy on that board, and every policy network must accept 61 inputs and return 14 outputs. Pollen’s FAQ in its GitHub repository says the robot’s neural accelerator “takes a small detector and little more,” and advises running anything larger off the robot.
Training is where the compute goes. Huang told the ducks they trained “in the cloud on a whole bunch of NVIDIA supercomputers.” The one training run the technical post describes used a single NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU, in a step the post calls “reinforcement learning in the cloud.” NVIDIA’s tutorial cites up to 1.25 million simulation steps per second when training across 16 GPUs, while Pollen’s README estimates one to two hours for a usable walking gait at 4,096 environments on a CUDA GPU it does not specify. RTX Spark, in NVIDIA’s account, runs the agents and replays finished policies in simulation.
One stack, two companies, one pending deal
Commercially, the pieces already fit. NVIDIA’s tutorial targets the same robot and the same ONNX interface as Pollen’s training code, and on September 3 NVIDIA announced that it had agreed to buy Hugging Face for US$12.93 billion. The deal has not closed. NVIDIA’s filing with the US Securities and Exchange Commission dates the agreement September 2, splits the price into about US11.9billionforshareholdersanduptoUS1.0 billion for employee retention, and expects closing in the first half of 2027, subject to regulatory approvals. TechCrunch’s keynote coverage says NVIDIA “bought Hugging Face in September,” which makes Pollen “now an Nvidia company”; the filing does not support that yet.
Huang’s announcement says “NVIDIA compute will not be required to build on or deploy through Hugging Face.” Pollen’s training code asks for a CUDA GPU but also offers an option to run training on Hugging Face Jobs instead. Open has a boundary here too: the press kit says the open-source statement covers the software stack, not the mechanical and electronic design files.
Preorders are not deliveries
Pollen’s launch post targeted first deliveries before Christmas 2026 in North America, Europe and the United Kingdom. As of October 26, its store carries a banner that reads: “The community ordered a lot of ducks. We can’t promise Christmas delivery for new microduck orders anymore, but we’re ramping up production. First ordered, first shipped.” It estimates four to six months for new orders and does not say how many earlier preorders will arrive by December 25.
The 20,000 figure is a preorder count, not a sales figure, from a company executive’s tweet relayed by TechCrunch. Pollen’s launch post, press kit and store page publish no production volume. For scale, Pollen says more than 10,000 Reachy Mini robots have reached customers since that product launched in 2025.
What would settle the question
The stronger result of the week may be how checkable the claims are. Pollen’s repositories, its published policies and NVIDIA’s tutorial are all public, and as of October 26, 503 model repositories outside Pollen’s own account carry “microduck” in the name. The test that would settle the agent question is a behavior nobody has published, with the reward bugs and actuator mismatches found by the agent itself, running on a robot a customer owns. Customer deliveries are the point where that test can leave the keynote stage.
The post The Microduck Demo Shows AI Agents Planning Robot Training, Not Solving Sim-to-Real appeared first on DataFLOQ.
