LIBERO offline: TinyVLA v1 (2.5GB) ratio 0.092 vs pi0.5-libero-finetuned (13.8GB) ratio 0.057; GR00T not loadable

Upload tinyvla_b200/renders_dagger_closedloop/cl_vla_stand.mp4 with huggingface_hub

Upload tinyvla_b200/renders_dagger_closedloop/cl_vla_ground_pick.mp4 with huggingface_hub

Upload tinyvla_b200/renders_dagger_closedloop/cl_vla_walking.mp4 with huggingface_hub

Upload tinyvla_b200/renders_dagger_closedloop/cl_vla_sitstand.mp4 with huggingface_hub

Upload tinyvla_b200/ckpt_dagger_iter0/README.md with huggingface_hub

DAgger iter0 checkpoint: first closed-loop STABLE behavior (sitstand 8s), loss 0.001-0.004

Upload tinyvla_b200/renders_closedloop/cl_vla_stand.mp4 with huggingface_hub

Upload tinyvla_b200/renders_closedloop/cl_vla_walking.mp4 with huggingface_hub

Upload tinyvla_b200/renders_closedloop/cl_teacher_walking.mp4 with huggingface_hub

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)

duck FT eval (fixed expert): 13/21 better than zero-baseline

sim renders: GT vs duck-FT policy (FIXED expert), external camera

microduck fine-tune ckpt (8k steps on duck v2 data), expert keys fixed

fix expert keys (_orig_mod stripped)

fix expert keys (_orig_mod stripped): prior file had a random flow head on load

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)

zero-shot LeKiwi eval of final v4 ckpt: ratio 1.64 all conditions (demos and descriptor both inert)

FINAL v4-allchannels checkpoint @ 60k global steps (11.5M samples, 163-dataset physical-ai stream)

microduck sim renders, external MuJoCo camera: GT teacher vs v4@20k side by side, offset-from-STAND action convention

v4-allchannels intermediate checkpoint @ ~20k global steps (3.8M samples over 163-dataset physical-ai stream) + README

fix: HubEpisodeStream.set_epoch was missing — first worker-recycle epoch boundary crashed the run

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)

streaming: episodes_per_epoch worker lifetime cap (dataloader workers leaked to OOM-kill on multi-hour runs twice)

microduck zero-shot renders + metrics (v4 ckpt step_10000): all combos worse than zero-baseline, honest negative

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

v4 config: shrink shuffle buffer/workers — 5 images per sample (2 cams + K=3 demos) made 2048-buffer x12 workers OOM the dataloader

specs: morph_key per dataset (6 robot families)

GR00T family robot text prompts

GR00T family morphology descriptors (estimates, flagged)

config: v4 all-channels on full physical-ai pool (smoke-verified, running)

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

streaming: in-context demo support cache (per-dataset episode ring, LRU), morphology+robot-text per sample

B200.md: measured cost-per-sample table across GPUs (3090/4090/H100/B200), consumer-card build pitfalls

guard package init: keep bench/data paths importable without lerobot

B200.md: measured smoke results (1xH100 495 s/s full script, 2xH100 DDP functional but interconnect-bound 340 s/s), spawn requirement, honest 8x caveat

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

streaming: stats fallback via episode-parquet sampling for repos without meta stats (Tuned-Tasks)

config: same yaml float fix

config: peak_flops in yaml-valid scientific notation

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

B200.md: install order with all caught pitfalls (hf CLI patterns, py3.12, torchvision pair, causal-conv1d isolation)

requirements: py3.12, torch/tv pair, lerobot from git tag, triton>=3.7.1 (fla hopper bwd bug)

add data __init__

add modules __init__

add modules/embodiment.py

add package __init__