VLA closed-loop driver (fixed head camera + auto-gain): falls <1s all behaviors — BC covariate shift on unstable system

closed-loop MicroDuck harness: obs.rs 61D builder validated by ONNX teacher (stands 8s @10Hz replan)

CRITICAL fix: torch.compile'd expert saved as expert._orig_mod.* -> from_pretrained silently dropped it (random flow head at eval). Save the uncompiled module

lekiwi zero-shot eval harness (hub streaming, 2x2 demo/descriptor grid)

eval_microduck: zero-shot eval+render harness (v4 mode: demos from same behavior, duck descriptor)

train_fast: non-persistent hub workers (recycle each epoch to reset leaks)

wds mixture: unitree+nav converters (gnm/vamos), wds source branch in train_fast, AMP autoselect (V100 fp16), embedding expand 16->32, push/fetch/validate tooling, DATA.md

train_fast: v4 all-channels via hub stream (support passthrough, morph_text tokenization, guarded shard morph block)

fix DDP hang: spawn dataloader workers (fork-after-CUDA/NCCL starved torchrun heartbeat), picklable Collate, fewer workers per rank

fix: yaml 1.1 parses '2250.0e12' as string -> float() cast in MFU calc (crashed first log step)

train_fast: torchrun DDP (per-rank streaming, rank0 logging/saves, global throughput)

tinyvla_b200: B200 streaming training on nvidia/physical-ai (745 samples/s e2e, 7x ViT fast path, hub episode streaming, full instructions)