license: apache-2.0
pretty_name: Shadow Dance v2
task_categories:
- reinforcement-learning
tags:
- robotics
- unitree-g1
- humanoid
- motion-control
- gear-sonic
Shadow Dance v2
Fifty-four team-authored synthetic Unitree G1 motion references for the Ultimate Bots SuperSONIC challenge: 12 parameterized partnerless dip training variants, 12 Shadow Gancho training variants, 10 conservative rehearsal motions (including forward walk and true heading turn in both directions), eight selection-validation motions, four disclosed preflight-only legacy v1 tests, and eight fresh untouched final-test motions.
The archive includes SONIC motion-lib PKLs, transparent source CSVs, fixed split lists, the full provenance/hash manifest, and the frozen MuJoCo validation report. No BONES-SEED motion, third-party video, or human biometric data is included.
Reference QA at the published revision: 54/54 pass, zero warnings, no joint-limit violation or MuJoCo self-contact, worst support/foot IK residual 6.96 mm / 3.74 degrees, and a positive 1.73 cm minimum hero support margin during the deepest hold. Gancho feet rise 29.4–31.4 cm. See the validation JSON for every metric and the GitHub repository for generation code and limitations.
Repository: https://github.com/Durp06/shadow-dance
The 30 v1 sequences and all 60 CSV/PKL payloads remain byte-identical to the immutable
source release:
https://github.com/cristpierce/shadow-dance/releases/tag/shadow-dip-v1.0.0
(archive SHA-256
94099f031b8a0b5ea36c809e705f77088342a6b54d73f9735508b146841c1370).
These are desired kinematic trajectories, not policy rollouts and not proof of physical robot execution.
Challenge acknowledgement requested by the Ultimate Bots portal: Motion Data by Bones Studio. No BONES-SEED motion or derived data is included in this archive.