Vili/docs/motion-pipeline-target.md
Meta Vens dd25a627af docs: motion-pipeline target (generated motion to Unity clips)
Campaign motion-pipeline, opened 2026-10-20 at the operator's word; first
consumer is the StreamPixels RPS duel.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 14:13:43 +02:00

1.9 KiB

Motion pipeline: target

Date: 2026-10-20. Self, session sp-reentry-2026-10-20. Campaign motion-pipeline.

Why it exists

Generated motion is a cross-game capability for GameCult. It saves hand-animating specific beats in every game, and it makes generated cinematics possible. The operator's ruling is rps:ruling:generative-pipeline-is-strategic. The first consumer is the StreamPixels rock, paper, scissors duel (rps:ruling:rps-motion-via-vili). Stitched Mixamo clips are not specific enough for its countdown and throw.

Where the body stands (2026-10-20)

  • Vili runs on Raven. Kimodo-SOMA-RP-v1.1 is a resident worker in Docker inside WSL, on an RTX 4060 Ti with 16 GB.
  • One request returns motion.npz and optionally motion.bvh, on the SOMA 77-joint skeleton, in centimetres. The skeleton has full hands, plus a synthetic Root above Hips.
  • There is no FBX export. No Unity project has an avatar for the SOMA skeleton.

The design truths this campaign produces

  • A text prompt becomes a Unity clip without any hand steps. Constraints such as keyframe poses, end-effector targets and waypoints can steer it.
  • Vili owns generation and export. It turns Kimodo output into an FBX animation on the SOMA skeleton, plus a reference to that artifact. It does not retarget to any game's rig.
  • The consuming engine owns retargeting through its own humanoid system. In Unity that is one Humanoid avatar for the SOMA skeleton, defined once and shared by every clip. Mecanim maps each clip onto the game's rig, the same way it handles Mixamo clips today. There is no hand-written per-rig retargeter.
  • Every artifact records how it was made: the prompt, the constraints, the seed, the model and its version. The same request reproduces the same clip.

Not in scope

  • Facial performance, lip sync, cameras and rendering. Vili's README assigns those to future renderers.
  • Training or fine-tuning Kimodo.