Oh sh*t, this $399 robot duck fell on its back and got up with no human touching it
And nobody programmed that move. It learned it inside a simulation, by falling and trying again, then the same policy was moved onto the real robot.
Left side is the simulation, right side is the real Microduck from Pollen Robotics (Hugging Face). Watch the dashboard react to every moment:
• Fall detection: the second it hits the floor, the screen flashes FALL DETECTED and the whole system flips into recovery mode
• Self recovery: it rolls, pulls its legs under its body, pushes up and stands. In slow motion you can see every step of the get-up
• Object contact: it walks up to a plush toy, lowers its center of mass and handles it without tipping over
• Sim to real: the simulated twin does the same moves at the same time. One brain, two bodies
We used to program robots line by line.
Now we let them fail in a simulation until they figure it out, then copy that brain into a real body.
And this one is open source. People are already training their own walks at home on one GPU.
Footage and links in the reply ↓
@pparik__ The token point is the real story. Most companies are paying for AI that still needs a human to check every output. Do you think that gap closes in 2027 or gets worse?
Oh sh*t, this $399 robot duck fell on its back and got up with no human touching it
And nobody programmed that move. It learned it inside a simulation, by falling and trying again, then the same policy was moved onto the real robot.
Left side is the simulation, right side is the real Microduck from Pollen Robotics (Hugging Face). Watch the dashboard react to every moment:
• Fall detection: the second it hits the floor, the screen flashes FALL DETECTED and the whole system flips into recovery mode
• Self recovery: it rolls, pulls its legs under its body, pushes up and stands. In slow motion you can see every step of the get-up
• Object contact: it walks up to a plush toy, lowers its center of mass and handles it without tipping over
• Sim to real: the simulated twin does the same moves at the same time. One brain, two bodies
We used to program robots line by line.
Now we let them fail in a simulation until they figure it out, then copy that brain into a real body.
And this one is open source. People are already training their own walks at home on one GPU.
Footage and links in the reply ↓
OH SH*T, HOW IS THIS POSSIBLE?
this $399 robot duck learned to kick, fall and get back up inside a simulation
then the exact same policy ran on the real robot. left is sim, right is real
it's open source. you can train your own walks at home
When @jensenhuang asks for a somersault, you deliver. 🦆
See how AI agents on NVIDIA RTX Spark help bring @huggingface's Microduck from simulation to the real world.
OH SH*T, HOW IS THIS POSSIBLE?
this $399 robot duck learned to kick, fall and get back up inside a simulation
then the exact same policy ran on the real robot. left is sim, right is real
it's open source. you can train your own walks at home
I GAVE KIMI K3.1 PRO AND OPUS 5.5 THE SAME PROMPT
one prompt, one 3D lighthouse scene running in the browser
top: Kimi K3.1 Pro
bottom: Opus 5.5
Kimi went for rock detail, foam and clouds. Opus went for light, gulls and lightning
which one would you ship?
I GAVE KIMI K3.1 PRO AND OPUS 5.5 THE SAME PROMPT
one prompt, one 3D lighthouse scene running in the browser
top: Kimi K3.1 Pro
bottom: Opus 5.5
Kimi went for rock detail, foam and clouds. Opus went for light, gulls and lightning
which one would you ship?