pi0.5-small (from scratch) — dual-xArm sim packing
Small pi0.5-style VLA trained from scratch on the dual-xArm simulation packing data,
as a sanity check / comparison against pi0.5 fine-tuning (openpi Pi0History, h120, pi05=True).
- VLM backbone: 12 layers, width 1024, mlp 4096 (random init)
- Action expert: 12 layers, width 512, mlp 2048 (random init)
- Shared attention: 12 heads x head_dim 128, MQA
- Vision: pretrained SigLIP So400m/14 trunk (from pi05_base), fully fine-tuned at 10x lower LR
- 25 epochs @ global batch 256 on 4x GH200; cosine 1e-4 -> 1e-5 (VLM/expert), 1e-5 -> 1e-6 (SigLIP)
Layout: adv/<step>/ and baseline/<step>/ contain orbax checkpoints
(params, train_state, assets) at ~25k-step (~5-epoch) milestones plus the final step.
Configs: pi05_small_scratch_xarm_sim_{adv,baseline}_history_h120 in
the small-scratch-pi05 branch of openpi-test.
