chennana1028/gwp05-openarm-sbint-abs-step50000

GigaWorld-Policy-0.5 · SBInt OpenArm (step 50000, absolute)

Fine-tuned GigaWorld-Policy-0.5 MoT transformer on SBInt openarm002 (873 episodes, 16-dim absolute joints, left-first packing), continued from step 20000 to 50000.

Sibling of the delta ckpt chennana1028/gwp05-openarm-sbint-delta-step50000. Same data / T5 / cameras / dims; only the action semantics differ:

• All 16 dims: absolute joint targets — delta_mask = [False]*16
• Left-first packing: [larm7, lgrip, rarm7, rgrip], gripper dims 7 and 15

Use with --action-mode abs and the matching norm_stats_sbint_abs.json in this repo. Mixing the delta norm stats gives ~10x scale errors (the scripts warn but do not abort).

Contents

File	Notes
config.json	CasualWorldActionTransformer_MoT, in/out_action_channels=16
diffusion_pytorch_model.bin	EMA weights, train job 4335486 step 50000 (continued from 4306868 step 20000)
norm_stats_sbint_abs.json	Matching absolute-joint quantile norm stats

Training snapshot

• Jobs: Slurm 4306868 (0-20k) + 4335486 (20k-50k) on 002-partition-RAD, 2x8 H100
• Steps: 50000 · ckpt every 1000 · eff batch 128
• Data: SBInt openarm002, LeRobot v3 (873 ep), task: place object in box + close lid
• Base: HF GWP-0.5 + Wan-class Diffusers VAE (lingbot_va_base)

Open-loop eval (episode 740, replan 48)

ckpt	MAE	seam ratio
abs step 50000 (this repo)	0.0130	3.67x
delta step 50000	0.0101	2.25x