π0.5 pre-training — landscape, throughput & π-equivalent sizing
estimateSizing π0-equivalent pre-training (match timesteps-seen) on our 8×H200 nodes — a reference-driven VLA landscape table, throughput/memory vs batch, and 1-node vs multi-node time & cost.
π0 target 0.7–1.4B ts1 node 31–62 d4 nodes 8–15 dnode tput 269 fr/s
π0.5 wrist-only — full-FT baselines
reportWrist-only π0.5 difficulty is mostly a masking artifact: physically removing the third-person camera from attention nearly recovers the both-camera ceiling.
attention drop 94.2%zero-mask 27.4%both-cam 96.6%LoRA floor 5.0%
π0.5 — architecture & tensor IO
model cardDual-expert flow-matching VLA, pinned down: per-module tensor IO for pi05_libero, the 3-camera input contract, and every ablation knob mapped to the exact module it touches.