A two-pound robot duck ran 4,096 parallel practice sessions inside a computer before ever touching a floor
A two pound robot duck ran 4,096 parallel practice sessions inside a computer before ever touching a floor, and that virtual run compressed years of practice into hours
This report details the Microduck robot demonstration at Nvidia's GTC Berlin, where a sub-800g unit trained in NVIDIA Isaac Sim using 4,096 parallel virtual environments to learn locomotion and manipulation skills. While the simulation pipeline compressed years of practice into hours by running up to 1,000x real-time speed on GPU workstations, the physical robot still exhibited instability during the live demo, with three of four units falling immediately. The piece highlights that while this 'simulation-first' approach reduced hardware development cycles from 18-24 months to seven for a separate UK humanoid startup in a German factory trial, transferring simulated skills to real hardware remains an unsolved challenge requiring significant tuning for gravity, friction, and uneven surfaces.
The article describes a demonstration of the Microduck robot trained using NVIDIA Isaac Sim with 4,096 parallel environments to compress learning time, noting that physical deployment still requires real-world tuning and lacks a confirmed timeline for widespread rollout.
Source: Robotics — Robot learning news discovery · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
parallel practice sessions
4,096
View original evidence
A two pound robot duck ran 4,096 parallel practice sessions inside a computer before ever touching a floor, and that virtual run compressed years of practice into hours
Open source S1Robot trajectories
780,000 trajectories
Trajectory count
View original evidence
aid. Separate published workflows from the chipmaker show a related number worth sitting with. Using the GR00T Blueprint, NVIDIA Isaac Sim trains robots at up to 1,000x real-time speed and GR00T Blueprint turns a few demos into 780K trajectories in 11 hours , equivalent to roughly 6,500 hours of human demonstration data. That ratio is the real engine behind what looks like fast learning on stage. The company has cont
Open source S10Reported duration
11 hours
View original evidence
aid. Separate published workflows from the chipmaker show a related number worth sitting with. Using the GR00T Blueprint, NVIDIA Isaac Sim trains robots at up to 1,000x real-time speed and GR00T Blueprint turns a few demos into 780K trajectories in 11 hours , equivalent to roughly 6,500 hours of human demonstration data. That ratio is the real engine behind what looks like fast learning on stage. The company has cont
Open source S10Reported duration
6,500 hours
View original evidence
aid. Separate published workflows from the chipmaker show a related number worth sitting with. Using the GR00T Blueprint, NVIDIA Isaac Sim trains robots at up to 1,000x real-time speed and GR00T Blueprint turns a few demos into 780K trajectories in 11 hours , equivalent to roughly 6,500 hours of human demonstration data. That ratio is the real engine behind what looks like fast learning on stage. The company has cont
Open source S10months
7
View original evidence
language alone. The Siemens trial, which ran for two weeks in January 2026 ahead of the April announcement, was the most demanding deployment to date. The simulation-first development approach, training and validating the robot’s behaviors in a virtual environment before physical deployment, allowed the startup Humanoid to slash hardware development from the usual 18, 24 months to just seven months , the companies s
Open source S9months
18
View original evidence
language alone. The Siemens trial, which ran for two weeks in January 2026 ahead of the April announcement, was the most demanding deployment to date. The simulation-first development approach, training and validating the robot’s behaviors in a virtual environment before physical deployment, allowed the startup Humanoid to slash hardware development from the usual 18, 24 months to just seven months , the companies s
Open source S9months
24
View original evidence
language alone. The Siemens trial, which ran for two weeks in January 2026 ahead of the April announcement, was the most demanding deployment to date. The simulation-first development approach, training and validating the robot’s behaviors in a virtual environment before physical deployment, allowed the startup Humanoid to slash hardware development from the usual 18, 24 months to just seven months , the companies s
Open source S9Reported duration
8 hours
View original evidence
ot in front of the audience. That same pipeline already runs at industrial scale elsewhere. A wheeled humanoid robot built by a UK robotics startup and powered by the same simulation stack recently completed an eight hour shift moving totes inside a real electronics factory in Germany. That trial mattered because it moved the idea past a stage prop. The evidence behind the speed claims None of this rests on marketing
Open source S8
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
A two pound robot duck ran 4,096 parallel practice sessions inside a computer before ever touching a floor, and that virtual run compressed years of practice into hours
Open source S1
aid. Separate published workflows from the chipmaker show a related number worth sitting with. Using the GR00T Blueprint, NVIDIA Isaac Sim trains robots at up to 1,000x real-time speed and GR00T Blueprint turns a few demos into 780K trajectories in 11 hours , equivalent to roughly 6,500 hours of human demonstration data. That ratio is the real engine behind what looks like fast learning on stage. The company has cont
Open source S10
language alone. The Siemens trial, which ran for two weeks in January 2026 ahead of the April announcement, was the most demanding deployment to date. The simulation-first development approach, training and validating the robot’s behaviors in a virtual environment before physical deployment, allowed the startup Humanoid to slash hardware development from the usual 18, 24 months to just seven months , the companies s
Open source S9
ot in front of the audience. That same pipeline already runs at industrial scale elsewhere. A wheeled humanoid robot built by a UK robotics startup and powered by the same simulation stack recently completed an eight hour shift moving totes inside a real electronics factory in Germany. That trial mattered because it moved the idea past a stage prop. The evidence behind the speed claims None of this rests on marketing
Open source S8
Source：Robotics — Robot learning news discovery · gcn.com