Downloads · 30 days
5
9% of all-time downloads
zhoumiaosen/pi0-libero10-full-finetune
pi0-libero10-full-finetune is a robotics model from zhoumiaosen. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot. The card lists the license as apache-2.0.
A fine-tune of lerobot/pi0base (Physical Intelligence's π₀, PaliGemma-2B + 300M flow-matching action expert) via LeRobot on the full LIBERO-Long (libero10) subset of HuggingFaceVLA/libero — the same 379 episodes, epis…
Downloads · 30 days
5
9% of all-time downloads
All-time downloads
53
Public
Parameters
4B
8.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors8.9 GB · 100%
How the weights are stored.
BF163.6B · 90%
From the Hugging Face model README
A fine-tune of lerobot/pi0_base
(Physical Intelligence's π₀, PaliGemma-2B + 300M flow-matching action expert)
via LeRobot on the full LIBERO-Long
(libero_10) subset of HuggingFaceVLA/libero — the same 379 episodes,
episode split, and 30,000-step budget as our companion
GR00T
and SmolVLA
fine-tunes, for a directly comparable result.
lerobot/pi0_baseHuggingFaceVLA/libero, episodes 0–378 (the 379 episodes
belonging to the 10 libero_10 tasks).--policy.train_expert_only=true: only the ~300M-parameter action expert
was trained; the ~3.7B-parameter PaliGemma VLM backbone was frozen.
Why: a full fine-tune of this model (4,028,019,472 params — 4B total)
OOM'd on the very first optimizer step, even with --policy.dtype=bfloat16 --policy.gradient_checkpointing=true. AdamW keeps per-parameter momentum
and variance state, which alone exceeds 24GB for a model this size
regardless of activation-memory savings from gradient checkpointing.
train_expert_only=true drops the trainable parameter count to
578,036,768 (578M) and training then ran cleanly at ~10.7GB. This also
happens to match our GR00T fine-tune's actual effective training scope
reasonably closely: GR00T's own fine-tune only ever trained its ~751M-param
diffusion action head, with its 3B backbone frozen throughout (see the
lora_rank caveat in GR00T's model card) — so both checkpoints in this
comparison set are, in practice, "train the action-generation head, keep
the pretrained vision-language backbone frozen," even though that wasn't
originally the explicit intent for either.HuggingFaceVLA/libero's per-episode metadata (data/chunk_index +
data/file_index, used to decide which of the dataset's 377 parquet files
to download for a given episode request) does not match the dataset's
actual physical file layout. Requesting --dataset.episodes=[0..378]
correctly computes from that metadata that 25 specific files are needed —
but those 25 files, once fetched, physically contain only episodes 0–60,
not 0–378 (verified directly: the file the metadata claims holds episode
378 actually holds episodes 58–60). The repository has 377 total data
files; episode 378's real data lives much further into that range than the
broken metadata implies, and no differently-scoped request could fix this,
since the mapping itself is wrong. Fix: construct the dataset once with
no episode restriction at all, forcing a full, correct download of the
entire data/ directory (confirmed: 383 files, matching the dataset's
total_frames=273465 metadata exactly) — this bypasses the broken
per-episode file selection entirely, since there's no selection to get
wrong when everything is fetched. Once locally cached this way, a normal,
episode-scoped dataset construction works correctly on its own. This is a
much larger download than our 379-episode slice alone needs (~78GB vs.
~2GB), but was the only reliable way to get correct training data.
Evaluated on the full libero_10 (LIBERO-Long) suite, 5 episodes per task
(50 total), same protocol as our other checkpoints. Eval episodes use
different seeds than training.
Overall: 0/50 = 0% success — identical to zero-shot π0 (also 0/50). Despite the loss dropping substantially and consistently over the full 30,000 steps (a real learning signal, not a stalled/broken run), this did not translate into any task success in this eval protocol.
Task (libero_10 id) | Description | Fine-tuned, 30k steps (5 ep) | Zero-shot (5 ep) |
|---|---|---|---|
| 0 | put both the alphabet soup and the tomato sauce in the basket | 0/5 | 0/5 |
| 1 | put both the cream cheese box and the butter in the basket | 0/5 | 0/5 |
| 2 | turn on the stove and put the moka pot on it | 0/5 | 0/5 |
| 3 | put the black bowl in the bottom drawer of the cabinet and close it | 0/5 | 0/5 |
| 4 | put the white mug on the left plate and put the yellow and white mug on the right plate | 0/5 | 0/5 |
| 5 | pick up the book and place it in the back compartment of the caddy | 0/5 | 0/5 |
| 6 | put the white mug on the plate and put the chocolate pudding to the right of the plate | 0/5 | 0/5 |
| 7 | put both the alphabet soup and the cream cheese box in the basket | 0/5 | 0/5 |
| 8 | put both moka pots on the stove | 0/5 | 0/5 |
| 9 | put the yellow and white mug in the microwave and close it | 0/5 | 0/5 |
A plausible, honest explanation (not verified in detail): with the VLM backbone frozen, only the ~300M-parameter action expert can adapt to LIBERO's action distribution and visual scene — a substantially smaller adaptable capacity than GR00T's ~751M-parameter diffusion head (also trained with a frozen backbone, but roughly 2.5x more parameters) or SmolVLA's fully-trainable ~450M-parameter model. π0's action expert may simply need more steps, more data, or a less constrained training scope (e.g. also unfreezing the VLM, which would require a larger GPU or a memory-efficient optimizer such as 8-bit AdamW) to see the same kind of improvement GR00T and SmolVLA showed on this benchmark with only 379 training episodes.
train_expert_only=true — the PaliGemma-2B VLM backbone in
this checkpoint is unchanged from lerobot/pi0_base; only the action
expert weights differ.