Huy-Quang-Dao / instantTrack
git clone https://github.com/Huy-Quang-Dao/instantTrack.git
Trains a whole-body motion tracking policy for the Unitree G1 humanoid on LAFAN1 clips, built on mjlab and MuJoCo Warp.
Why you might care
Aims to reach usable G1 motion tracking across all 40 LAFAN1 clips in about 15 minutes on one RTX 5090, versus 0.55 to 1.5 hours for the SuperMimic baselines it compares against. Adds a policy EMA, GPU-side training loop changes, and a contact reward for cleaner foot contact. Benchmark numbers are self-reported from the repo's own runs.
+6
stars, 24h · #5,806
+76
stars, 7 days · #4,223
+76
stars, 30 days · #11,461
76
★ total · +76% this week
4
forks · 1 open issues
Star history (as recorded here)
LanguagePython
LicenseMIT
Created2026-10-29 (3d ago)
Last push1d ago
Watchers0
OwnerUser
Links
- GitHub: github.com/Huy-Quang-Dao/instantTrack
- Homepage: https://huy-quang-dao.github.io/instantTrack/
- Releases · Issues · Hacker News mentions
README
Loading README…
Blurb generated by claude-haiku-5-5 2d ago. Star deltas marked ~ are extrapolated from partial history.