Holo

Holo teaches a humanoid a new skill from one ordinary video. Show it any clip of a person (a kick, a dance step, a jump) and a physics-simulated humanoid robot performs it: no training, no motion-capture suit, no per-skill reinforcement learning. How it works: - See the body: SAM 3D Body reconstructs the person's full 3D body in every frame of a normal phone or YouTube video. - Understand the intent: Muse Spark watches the clip and says what the person is doing and why ("a side kick aimed at head height, weight on the left leg"), so the robot copies the goal, not just joint angles. - Retarget: the 3D joints are fitted to a humanoid body with inverse kinematics in MuJoCo. - Perform with physics: Meta Motivo, a behavioral foundation model, takes the motion as a prompt and controls the humanoid zero-shot, with real balance, gravity and contacts. - Talk to it: "same kick but slower", "mirror it", "kick lower". Muse turns the request into a new prompt for Motivo. - Robot-ready: the motion is shown on a robot body and exported for a real humanoid (Unitree G1 retarget). Why it's new: research systems that learn from video (e.g. VideoMimic, HumanX) train a robot policy for hours for each skill. Holo needs one video and no training, because Meta Motivo already knows how to move, SAM 3D Body sees the motion, and Muse Spark understands it. Status: the video-to-3D body pipeline (SAM 3D Body) and the Motivo physics humanoid are running. We are connecting retargeting and zero-shot tracking end to end, then adding Muse intent and language edits. Holo started as a live AI guide that sees what you see. Now it turns what it sees into something a robot can do.