Got to bring Jiji on stage at NVIDIA's RTX Spark launch event! 🐈‍⬛ When I started building her a year ago, I never imagined I'd be here. Getting to bring her on stage and share that journey felt really special. I barely slept the night before due to nervousness and excitement. Thank you to @NVIDIAKorea for inviting me to talk about Jiji and the dream I’m still working toward. Honored to be there, and so excited to keep building Jiji version 2. 🐾 @NVIDIARobotics @nvidia

I had a Lion King moment with Jiji at @NVIDIAKorea RTX Spark laptop looks cool :) @NVIDIARobotics @NVIDIAKorea

her walking gait is so cute hahahah this is a MicroDuck based cat robot I am building, and the thing is that I am using Robotis's XL330 motors and have placed Jetson Orin Nano inside of the head and it might be too heavy for the motors. Should I move the Jetson board to the body? Any suggestions would be appreciated! @pollenrobotics @huggingface @NVIDIARobotics @nvidia

jiji is standing!! I can’t wait to see her first walk 🥹

okay the deskbot is looking cute 😵‍💫 @grok @bot

why does it look like a hanster

I spent days trying to train my MicroDuck to stand up, but the RL policy kept getting stuck in a butt-up position. 🦆 Many people suggested changing the reward function or PPO parameters. But the breakthrough was in the training state distribution. Two things made the difference: 1. Use the actual failure state. Instead of spawning the robot at different body angles with joints at HOME, I captured the real stuck configuration and used its joint positions. The folded neck was part of the trap. 2. Train across both variables. Body tilt alone wasn't enough. The policy needed experience across the combination of body tilt + head/neck configuration, not root orientation alone. With that targeted curriculum, face-down recovery went from 0% → 100% without changing the reward function. Even more interesting: face-up recovery also went from 0% → 100%, despite never being directly targeted. The baseline face-up rollout ended in the same butt-up attractor, so learning to escape that state solved both failures. I’m having so much fun with this MicroDuck project. Every time I think I’ve found the problem, I end up learning something completely new about reinforcement learning ! *Microduck is trained on NVIDIA Jetson AGX Thor @pollenrobotics @seeedstudio @huggingface @NVIDIARobotics #microduck

I ran 4,096 simulated environments on NVIDIA Jetson Thor. When initialized face-down, the policy converges to a stable “butt-up” position instead of recovering to standing. This state receives about 68% of a proper standing pose. This video explains what I learnt about Reinforcement Learning and PPO algorithm. Microduck has provided a good platform to try out RL for the first time ! Excited to run more experiments and always let me know if any other suggestions :) @pollenrobotics @seeedstudio @huggingface @NVIDIARobotics #microduck

my ducks think butt up is standing up, any suggestions on how to fix this? @RemiFabreRobot @pollenrobotics @NVIDIARobotics #NVIDIA

SO-101 ARM DONE!! so proud of us 🙋🏻‍♀️🐈‍⬛🎀 I think it goes well with my cat robot, what do you think? 🤠 @seeedstudio @huggingface

I ran MicroDuck RL on a Jetson Thor, and poor duck is STRUGGLING 🐣 Getting it running on Thor was a bit of a rabbit hole. The original MicroDuck environment used PyTorch 2.9 + CUDA 12.9, but Thor’s GPU is sm_110, which that PyTorch build didn’t have a kernel for. So I had to switch to NVIDIA’s Jetson PyTorch build with CUDA 13, which supports Thor’s architecture. Then get it working alongside MuJoCo Warp + mjlab. Now it runs PPO directly on the Thor GPU. 10,200 steps in and this is what we’ve got 🥸 @ClementDelangue @huggingface @pollenrobotics @NVIDIARobotics #NVIDIA

bullying this duck until you send me one @pollenrobotics @ClementDelangue

❤️❤️❤️❤️ my current favorite PCBs

I just built a 2-stage planetary gearbox for a NEMA 17 motor! 🪐⚙️ A planetary gearbox uses a sun gear, planet gears, a ring gear, and a carrier to achieve high gear reduction in a compact package. This one uses two 5:1 planetary stages stacked together. So here's the question: how many times would the output torque ideally increase compared to the original motor? 👀 Design by Domino on Maker World: makerworld.com/en/models/277… #robotics #gear #cyberdeck #engineeringexplained #gearbox

🫪

spoiler for my next 2 videos 🙈 can anyone guess the two projects I am currently working on? 👀

I designed custom flex PCB for my tamagotchi cyberdeck! The main challenge was to design it in the way to fit inside the old Sony Ericsson phone and align with existing keypad arrays, and it turned out pretty well! :)) #cyberdeck #tamagotchi #esp32 #cyberpunk #pcbdesign

planetary gears 😳