RLDX2 — PointWAM (Volt-H) on RB-Y1M + Wuji Hand 2
Code: PointWAM branch rldx2/frontier-deploy (experiment midtrain_rby1wuji_volt_h, PW_KP_PER_HAND=11).
| path | what |
|---|---|
midtrain/rby1wuji-midtrain-v1-kp11-bs64-dbg/model-step20000.pt |
Mid-train on RB-Y1M+Wuji2 teleop (790 ep / 8 h), init pt-flows-full-h, bs 64, 20k steps. Held-out robot_l2_moved 0.045. |
humanft/rby1wuji-humanft-vitra-jfo-kp11-bs64/model-step10000_frozenret.pt |
FT-A: human-only FT (frontier_human_261007, VITRA jumpfix_offset retarget), 10k steps, retargeter = mid-train (frozen). |
humanft/rby1wuji-humanft-vitra-jfo-cap30-kp11-bs64/model-step10000_frozenret.pt |
FT-A cap30: same as FT-A, CoTracker tracks with the 30 cm / 0.1 s physical cap (point-level visibility only). |
humanft/rby1wuji-humanft-rrc-nogate-kp11-bs64/model-step10000_frozenret.pt |
FT-B: human-only FT, RRC optimiser retarget, 10k steps, frozen retargeter. |
stats/rby1wuji-midtrain-v1/norm_stats.json |
Norm stats used by all (FT did not recompute stats). |
rldx2_bundle/ |
Robot-PC deploy bundle (env, code, weights, FFS, sample data, ckpt_<run>.tar per checkpoint, install.sh). |
Checkpoints hold model and model_ema (use EMA at deploy), no optimizer state. Load with
torch.load(..., weights_only=False); frozen text encoder / Mosaic3D tensors are rebuilt from their own source at load.
Overrides: experiment=midtrain_rby1wuji_volt_h data.norm_stats_path=stats/rby1wuji-midtrain-v1 + env PW_KP_PER_HAND=11.
Training recipe: constant LR 1e-4, AdamW wd 0.01, EMA 0.9999, 4 GPU × bs 16.