Yaman｜達人の動きをロボへ｜Physical AI
@yaman_xr
イマクリエイトCEO｜"体験を、拡張せよ"｜XR→Physical AI→BMI｜今は達人の動きをロボに移植（けん玉ロボ→溶接・左官・塗装）｜WeldNext(溶接VR・海外展開)｜すごいベンチャー100 2024・東大IPC他出資｜住商→トルコでクーデター遭遇→起業｜マトリックス着想｜採用中⬇️
参加 December 2009
1K フォロー中 1.9K ファン
これ、すごい！
けん玉の手をメインで移植してたけど（3日目にG1ちゃんがびっくり腰になったこともあり）、これができるなら、全然足を使ったいろんな技もいけるね！
もっと見る
The Unitree G1 executed kicks forward jumps and squats after ASAP trained policies from retargeted human motion and corrected the sim to real gap with rollout data from the physical robot.
The key detail is the hardware loop. The team measured how the real G1 responded trained a delta action model from that data and used it to reduce the mismatch between MuJoCo and the robot.
This matters because humanoid control often breaks when the foot hits the floor.
A jump or fast leg swing can expose errors in timing friction torque and balance.
ASAP shows how learned motion can move from simulation onto real hardware with less collapse at contact.
もっと見る
