NVIDIA Kimodo generates human body motion from text. Give 1 to 10 prompts, each starting with "A person" and describing one or two actions, plus a duration of 1 to 10 s per prompt; the moves play in order with blended transitions. An optional JSON constraints file sets a ground path (root2d waypoints) or keyframes. Output: a BVH at 30 fps on the 77-joint SOMA skeleton (centimeters, Y-up) and the raw NPZ, ready to retarget. 1 to 4 variations per run. No mesh or character is included.