π0.5 reports

A dual-expert flow-matching VLA, pinned down: every module's input/output tensor shape for pi05_libero, the camera-input contract, and each GitHub-issue ablation knob mapped to the exact module it touches.

• Dual-expert Gemma (PaliGemma gemma_2b 2048-w + action expert gemma_300m 1024-w), 18 layers, joint attention
• SigLIP So400m/14 image tower → 256 tokens/camera; 3 fixed image slots (base + left/right wrist)
• Per-module tensor IO table + flow-matching train/sample loop
• Ablation knobs: action_horizon, action_dim padding, LoRA vs full-FT, head pretraining, view masking, fps/codec, multi- vs per-task

The apparent difficulty of wrist-only π0.5 was largely an artifact of the masking mechanism, not a capacity ceiling: physically removing the third-person camera from attention nearly recovers the both-camera ceiling.

• 5 full-FT baselines (2 evaluated, 3 in-flight self-distillation ablations)
• How each is trained & wired into the openpi pipeline, with the line of code per mechanism
• GitHub issues #5 (wrist-only floor) · #10 (curriculum + distillation)