We are excited to share the latest progress on our open humanoid robot, ROBOTIS AI Sapiens!
In this new video, our team demonstrates how AI Sapiens learns complex movements directly from raw 2D video data without any specialized motion capture equipment.
By extracting 3D joint movements from a simple video and combining it with reinforcement learning in NVIDIA Isaac Sim, we successfully trained the robot to mimic fluid dance choreography.
One of the biggest challenges in humanoid robotics is the Sim2Real gap. To overcome this hurdle, we utilized our DYNAMIXEL-Q actuators. Thanks to their high precision and responsiveness, we managed to achieve an ultra-low Sim2Real gap, perfectly translating simulated joint movements into physical real-world control.
Check out the full video to see how this fluid choreography was made:
https://youtu.be/NeL2BmgcFJs