Retargeting

SkillDev tools

Use when working on Workbench motion retargeting, SONIC retargeted motion artifacts, SkyPilot retargeting YAMLs, or retargeting CLI behavior.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Retargeting skill

What this skill tells your AI

The instructions your AI receives, as published by nebius/nebius-physical-ai in skills/tools/retargeting/SKILL.md and read by ahel’s review.

Retargeting converts already-retargeted SOMA/G1/Bones motion artifacts into the real motion-lib PKL schema consumed by SONIC locomotion training.

Interfaces

CLI:

npa workbench sonic retargeting run
npa workbench sonic retargeting workflow
npa workbench sonic retargeting status
npa workbench sonic retargeting list

SkyPilot YAML:

  • workflows/testing/retargeting.yaml
  • workflows/testing/sonic-locomotion-finetuning.yaml

Routing And Data Flow

Retargeting is CPU-only by default.

Inputs and outputs use S3 paths:

  • --input-path: source motion prefix or object.
  • --output-path: retargeted motion output prefix.
  • --retarget-map: optional map artifact.

The result artifacts are real .pkl motion-lib files plus retargeting_result.json metadata. Do not replace this with a manifest-only shim.

Raw BVH inputs can be converted to upstream SONIC SOMA skeleton PKLs with extract_soma_joints_from_bvh.py, but upstream SONIC does not bundle a raw BVH-to-G1 robot retargeter. Use external SOMA Retargeter/GMR before the final SONIC motion-lib conversion when starting from raw BVH.

Workflow Constraint

Keep orchestration logic in SkyPilot YAML. Do not add a Python runner script for the SONIC locomotion fine-tuning path.

Signals

GitHub stars
28
Forks
15
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
retargeting
Source
github.com/nebius/nebius-physical-ai