Base Model: IDM-OMX-v2 (20K steps checkpoint)
Dataset: 558 episodes, 257,466 frames (OMX combined + SO-101 megamix FK→IK converted)
Steps: 50,000 (6.21 epochs)
Batch Size: 16 (effective 32 with gradient accumulation)
Learning Rate: 1e-4 (cosine schedule)
Duration: 5h 08min on RTX 5090
Final Train Loss: 0.2644

Evaluation

Metric	Baseline (10K)	v2 (20K)	v3 (50K)
Normalized MAE	0.6092	0.5582	0.3551 (-36.4%)
Forward Loss	0.3942	0.3453	0.2517 (-27.1%)

Joint	v2	v3	Change
shoulder_pan	82.12	71.76	-12.6%
shoulder_lift	83.01	72.74	-12.4%
elbow_flex	65.65	30.12	-54.1%
wrist_flex	141.40	96.43	-31.8%
wrist_roll	67.87	56.43	-16.9%
gripper	6.50	4.21	-35.2%

Usage