NVIDIA’s GTC Berlin keynote clip of Hugging Face’s Microduck robotic invitations an enormous conclusion: that AI brokers took a US$399 biped from simulation to the true world. NVIDIA’s personal engineering write-up describes a narrower job, planning an impediment course, constructing its props and assembling it in Isaac Sim, and it places “detailed {hardware} deployment” outdoors its scope. That could be a actual step for agent tooling, however the recipe that makes a 25 cm robotic stroll on an actual flooring existed earlier than the keynote.
What NVIDIA confirmed on stage
On October 21, close to the top of his Berlin, Germany, keynote, NVIDIA chief govt Jensen Huang held up an RTX Spark, the PC platform NVIDIA launched in laptops on October 16, with as much as 128GB of unified reminiscence and 1 petaflop of AI compute. He then performed a video of an agent serving to practice Microduck. NVIDIA’s dwell recap says the agent arrange a course, simulated the robotic in Isaac Sim, despatched coaching jobs to the cloud and introduced the realized coverage again to RTX Spark for analysis “earlier than deploying it to the bodily robotic.” 4 Microducks then joined him on stage and sat, stood and napped on cue.
Huang instructed the geese, “you realized your abilities proper 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 ship.” The publish added that AI brokers on RTX Spark “assist deliver” Microduck “from simulation to the true world.” TechCrunch reported on October 23 that Hugging Face chief govt Clem Delangue mentioned in a tweet that preorders had handed 20,000 items.
What the brokers did, in NVIDIA’s personal phrases
Two days later, NVIDIA’s technical weblog laid out the workflow. A immediate goes to Hermes Agent, an open-source agent from Nous Analysis, and to NVIDIA’s Nemotron 3.5 Bolt mannequin, each working on RTX Spark. They flip the immediate right into a step-by-step plan for an impediment 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 robotic’s coverage. The skilled coverage returns to RTX Spark for analysis in simulation.
The coaching step builds on current work. The publish says the experiment used “the Isaac Lab Microduck instance surroundings as its place to begin,” with a locomotion coverage “pretrained in that surroundings” that then realized ramps and obstacles throughout 4,096 parallel environments. It additionally states that for every activity a developer defines “what the robotic observes, the actions it might take, and the rewards used to information studying.” The publish’s part on the bodily robotic carries a restrict: “Detailed {hardware} deployment is outdoors the scope.”
NVIDIA’s separate Hugging Face tutorial, printed October 21, covers the true robotic. It has a developer practice a strolling coverage with a typed command, export it as an ONNX file and duplicate it to the robotic’s onboard pc. The phrase “agent” doesn’t seem in it. The tutorial says Isaac Lab “already offers Microduck’s robotic mannequin, scene, observations, actions, rewards, and area randomization.”
The implication is that the brokers’ documented work sits on the entrance of the pipeline, in planning and asset creation, whereas the step that places a coverage on an actual robotic runs by a recipe somebody had already constructed. The keynote recap and the technical publish describe the identical workflow however declare totally different scopes, and the technical publish is the extra cautious of the 2.
A somersault that predates the keynote
The somersault has a public historical past. Pollen Robotics, the Bordeaux, France, robotics group that joined Hugging Face in April 2025, lists a “Roulade” activity in its coaching repository: a “ahead roll over the pinnacle, land again on the ft.” The commit historical past exhibits that work touchdown between August 11 and August 18. Pollen’s official coverage repository on Hugging Face has held a file named roulade.onnx since its first add on August 31, 4 days after preorders opened.
Neither NVIDIA’s recap nor its technical publish says which coverage the stage geese ran, in order that file proves nothing concerning the keynote itself. It does present the habits predates RTX Spark laptops, which launched on October 16. The duty’s docstring carries the label “try 3, run 2” and information that the primary run realized “a violent ballistic ‘breakdance’ whip,” which was “optimum beneath the run-1 rewards.” Among the many fixes, rotation now counts solely whereas the robotic touches the bottom.
The place sim-to-real nonetheless bites
Pollen’s README names the issue at this scale: “actuator constancy is many of the sim2real hole.” Tiny servos drive a robotic of beneath 800 grams, so the coaching simulation fashions every servo’s voltage management legislation, randomizes command delay and battery voltage sag, and, in its backlash variants, provides one diploma of substances play both means on each joint. NVIDIA’s technical publish leans on the identical repair, a Higher Actuator Mannequin, then provides area randomization as a result of “sensors, floor contact, and management timing can differ on the true robotic.”
The identical repository exhibits what that hole prices in observe. Its AGENTS.md file, within the repository since launch week, opens by saying that each conference in it “exists as a result of breaking it produced a coverage that labored within the viewer and failed on {hardware}.” A centre-of-mass randomizer that accrued throughout resets “degraded each future for months.” A goal standing peak unsuitable by 5 millimetres “turned the purpose into an inconceivable goal for days.”
The README says the information is “additionally geared toward AI coding brokers working on this repo,” and brokers already work there: 96 of the repository’s 1,255 commits carry a Claude co-author trailer. One studying is that brokers develop into helpful for sim-to-real as soon as folks have written down what breaks. NVIDIA’s technical publish doesn’t present an agent discovering a failure mode like these above, which by Pollen’s personal account can value days to months.
The heavy compute lives elsewhere
The duck’s personal pc is small. Pollen’s press equipment lists a Rockchip RK3566 processor, from China-based Rockchip, with 1GB of RAM, 32GB of storage and a detachable battery good for round one hour of runtime. A 50 Hz loop executes the coverage on that board, and each coverage community should settle for 61 inputs and return 14 outputs. Pollen’s FAQ in its GitHub repository says the robotic’s neural accelerator “takes a small detector and little extra,” and advises working something bigger off the robotic.
Coaching is the place the compute goes. Huang instructed the geese they skilled “within the cloud on a complete bunch of NVIDIA supercomputers.” The one coaching run the technical publish describes used a single NVIDIA RTX PRO 6000 Blackwell Workstation Version GPU, in a step the publish calls “reinforcement studying within the cloud.” NVIDIA’s tutorial cites as much as 1.25 million simulation steps per second when coaching throughout 16 GPUs, whereas Pollen’s README estimates one to 2 hours for a usable strolling gait at 4,096 environments on a CUDA GPU it doesn’t specify. RTX Spark, in NVIDIA’s account, runs the brokers and replays completed insurance policies in simulation.
One stack, two firms, one pending deal
Commercially, the items already match. NVIDIA’s tutorial targets the identical robotic and the identical ONNX interface as Pollen’s coaching code, and on September 3 NVIDIA introduced that it had agreed to purchase Hugging Face for US$12.93 billion. The deal has not closed. NVIDIA’s submitting with the US Securities and Alternate Fee dates the settlement September 2, splits the value into about US11.9billionforshareholdersanduptoUS1.0 billion for worker retention, and expects closing within the first half of 2027, topic to regulatory approvals. TechCrunch’s keynote protection says NVIDIA “purchased Hugging Face in September,” which makes Pollen “now an Nvidia firm”; the submitting doesn’t assist that but.
Huang’s announcement says “NVIDIA compute is not going to be required to construct on or deploy by Hugging Face.” Pollen’s coaching code asks for a CUDA GPU but in addition presents an choice to run coaching on Hugging Face Jobs as a substitute. Open has a boundary right here too: the press equipment says the open-source assertion covers the software program stack, not the mechanical and digital design information.
Preorders usually are not deliveries
Pollen’s launch publish focused first deliveries earlier than Christmas 2026 in North America, Europe and the UK. As of October 26, its retailer carries a banner that reads: “The neighborhood ordered loads of geese. We will’t promise Christmas supply for brand new microduck orders anymore, however we’re ramping up manufacturing. First ordered, first shipped.” It estimates 4 to 6 months for brand new orders and doesn’t say what number of earlier preorders will arrive by December 25.
The 20,000 determine is a preorder depend, not a gross sales determine, from an organization govt’s tweet relayed by TechCrunch. Pollen’s launch publish, press equipment and retailer web page publish no manufacturing quantity. For scale, Pollen says greater than 10,000 Reachy Mini robots have reached clients since that product launched in 2025.
What would settle the query
The stronger results of the week could also be how checkable the claims are. Pollen’s repositories, its printed insurance policies and NVIDIA’s tutorial are all public, and as of October 26, 503 mannequin repositories outdoors Pollen’s personal account carry “microduck” within the identify. The take a look at that may settle the agent query is a habits no person has printed, with the reward bugs and actuator mismatches discovered by the agent itself, working on a robotic a buyer owns. Buyer deliveries are the purpose the place that take a look at can depart the keynote stage.
NVIDIA’s GTC Berlin keynote clip of Hugging Face’s Microduck robotic invitations an enormous conclusion: that AI brokers took a US$399 biped from simulation to the true world. NVIDIA’s personal engineering write-up describes a narrower job, planning an impediment course, constructing its props and assembling it in Isaac Sim, and it places “detailed {hardware} deployment” outdoors its scope. That could be a actual step for agent tooling, however the recipe that makes a 25 cm robotic stroll on an actual flooring existed earlier than the keynote.
What NVIDIA confirmed on stage
On October 21, close to the top of his Berlin, Germany, keynote, NVIDIA chief govt Jensen Huang held up an RTX Spark, the PC platform NVIDIA launched in laptops on October 16, with as much as 128GB of unified reminiscence and 1 petaflop of AI compute. He then performed a video of an agent serving to practice Microduck. NVIDIA’s dwell recap says the agent arrange a course, simulated the robotic in Isaac Sim, despatched coaching jobs to the cloud and introduced the realized coverage again to RTX Spark for analysis “earlier than deploying it to the bodily robotic.” 4 Microducks then joined him on stage and sat, stood and napped on cue.
Huang instructed the geese, “you realized your abilities proper 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 ship.” The publish added that AI brokers on RTX Spark “assist deliver” Microduck “from simulation to the true world.” TechCrunch reported on October 23 that Hugging Face chief govt Clem Delangue mentioned in a tweet that preorders had handed 20,000 items.
What the brokers did, in NVIDIA’s personal phrases
Two days later, NVIDIA’s technical weblog laid out the workflow. A immediate goes to Hermes Agent, an open-source agent from Nous Analysis, and to NVIDIA’s Nemotron 3.5 Bolt mannequin, each working on RTX Spark. They flip the immediate right into a step-by-step plan for an impediment 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 robotic’s coverage. The skilled coverage returns to RTX Spark for analysis in simulation.
The coaching step builds on current work. The publish says the experiment used “the Isaac Lab Microduck instance surroundings as its place to begin,” with a locomotion coverage “pretrained in that surroundings” that then realized ramps and obstacles throughout 4,096 parallel environments. It additionally states that for every activity a developer defines “what the robotic observes, the actions it might take, and the rewards used to information studying.” The publish’s part on the bodily robotic carries a restrict: “Detailed {hardware} deployment is outdoors the scope.”
NVIDIA’s separate Hugging Face tutorial, printed October 21, covers the true robotic. It has a developer practice a strolling coverage with a typed command, export it as an ONNX file and duplicate it to the robotic’s onboard pc. The phrase “agent” doesn’t seem in it. The tutorial says Isaac Lab “already offers Microduck’s robotic mannequin, scene, observations, actions, rewards, and area randomization.”
The implication is that the brokers’ documented work sits on the entrance of the pipeline, in planning and asset creation, whereas the step that places a coverage on an actual robotic runs by a recipe somebody had already constructed. The keynote recap and the technical publish describe the identical workflow however declare totally different scopes, and the technical publish is the extra cautious of the 2.
A somersault that predates the keynote
The somersault has a public historical past. Pollen Robotics, the Bordeaux, France, robotics group that joined Hugging Face in April 2025, lists a “Roulade” activity in its coaching repository: a “ahead roll over the pinnacle, land again on the ft.” The commit historical past exhibits that work touchdown between August 11 and August 18. Pollen’s official coverage repository on Hugging Face has held a file named roulade.onnx since its first add on August 31, 4 days after preorders opened.
Neither NVIDIA’s recap nor its technical publish says which coverage the stage geese ran, in order that file proves nothing concerning the keynote itself. It does present the habits predates RTX Spark laptops, which launched on October 16. The duty’s docstring carries the label “try 3, run 2” and information that the primary run realized “a violent ballistic ‘breakdance’ whip,” which was “optimum beneath the run-1 rewards.” Among the many fixes, rotation now counts solely whereas the robotic touches the bottom.
The place sim-to-real nonetheless bites
Pollen’s README names the issue at this scale: “actuator constancy is many of the sim2real hole.” Tiny servos drive a robotic of beneath 800 grams, so the coaching simulation fashions every servo’s voltage management legislation, randomizes command delay and battery voltage sag, and, in its backlash variants, provides one diploma of substances play both means on each joint. NVIDIA’s technical publish leans on the identical repair, a Higher Actuator Mannequin, then provides area randomization as a result of “sensors, floor contact, and management timing can differ on the true robotic.”
The identical repository exhibits what that hole prices in observe. Its AGENTS.md file, within the repository since launch week, opens by saying that each conference in it “exists as a result of breaking it produced a coverage that labored within the viewer and failed on {hardware}.” A centre-of-mass randomizer that accrued throughout resets “degraded each future for months.” A goal standing peak unsuitable by 5 millimetres “turned the purpose into an inconceivable goal for days.”
The README says the information is “additionally geared toward AI coding brokers working on this repo,” and brokers already work there: 96 of the repository’s 1,255 commits carry a Claude co-author trailer. One studying is that brokers develop into helpful for sim-to-real as soon as folks have written down what breaks. NVIDIA’s technical publish doesn’t present an agent discovering a failure mode like these above, which by Pollen’s personal account can value days to months.
The heavy compute lives elsewhere
The duck’s personal pc is small. Pollen’s press equipment lists a Rockchip RK3566 processor, from China-based Rockchip, with 1GB of RAM, 32GB of storage and a detachable battery good for round one hour of runtime. A 50 Hz loop executes the coverage on that board, and each coverage community should settle for 61 inputs and return 14 outputs. Pollen’s FAQ in its GitHub repository says the robotic’s neural accelerator “takes a small detector and little extra,” and advises working something bigger off the robotic.
Coaching is the place the compute goes. Huang instructed the geese they skilled “within the cloud on a complete bunch of NVIDIA supercomputers.” The one coaching run the technical publish describes used a single NVIDIA RTX PRO 6000 Blackwell Workstation Version GPU, in a step the publish calls “reinforcement studying within the cloud.” NVIDIA’s tutorial cites as much as 1.25 million simulation steps per second when coaching throughout 16 GPUs, whereas Pollen’s README estimates one to 2 hours for a usable strolling gait at 4,096 environments on a CUDA GPU it doesn’t specify. RTX Spark, in NVIDIA’s account, runs the brokers and replays completed insurance policies in simulation.
One stack, two firms, one pending deal
Commercially, the items already match. NVIDIA’s tutorial targets the identical robotic and the identical ONNX interface as Pollen’s coaching code, and on September 3 NVIDIA introduced that it had agreed to purchase Hugging Face for US$12.93 billion. The deal has not closed. NVIDIA’s submitting with the US Securities and Alternate Fee dates the settlement September 2, splits the value into about US11.9billionforshareholdersanduptoUS1.0 billion for worker retention, and expects closing within the first half of 2027, topic to regulatory approvals. TechCrunch’s keynote protection says NVIDIA “purchased Hugging Face in September,” which makes Pollen “now an Nvidia firm”; the submitting doesn’t assist that but.
Huang’s announcement says “NVIDIA compute is not going to be required to construct on or deploy by Hugging Face.” Pollen’s coaching code asks for a CUDA GPU but in addition presents an choice to run coaching on Hugging Face Jobs as a substitute. Open has a boundary right here too: the press equipment says the open-source assertion covers the software program stack, not the mechanical and digital design information.
Preorders usually are not deliveries
Pollen’s launch publish focused first deliveries earlier than Christmas 2026 in North America, Europe and the UK. As of October 26, its retailer carries a banner that reads: “The neighborhood ordered loads of geese. We will’t promise Christmas supply for brand new microduck orders anymore, however we’re ramping up manufacturing. First ordered, first shipped.” It estimates 4 to 6 months for brand new orders and doesn’t say what number of earlier preorders will arrive by December 25.
The 20,000 determine is a preorder depend, not a gross sales determine, from an organization govt’s tweet relayed by TechCrunch. Pollen’s launch publish, press equipment and retailer web page publish no manufacturing quantity. For scale, Pollen says greater than 10,000 Reachy Mini robots have reached clients since that product launched in 2025.
What would settle the query
The stronger results of the week could also be how checkable the claims are. Pollen’s repositories, its printed insurance policies and NVIDIA’s tutorial are all public, and as of October 26, 503 mannequin repositories outdoors Pollen’s personal account carry “microduck” within the identify. The take a look at that may settle the agent query is a habits no person has printed, with the reward bugs and actuator mismatches discovered by the agent itself, working on a robotic a buyer owns. Buyer deliveries are the purpose the place that take a look at can depart the keynote stage.
