Downloads · 30 days
15
19% of all-time downloads
zhoumiaosen/groot-n1p5-lora-libero10-full-finetune
groot-n1p5-lora-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 real fine-tune of NVIDIA GR00T-N1.5-3B via LeRobot's groot policy wrapper on the full LIBERO-Long (libero10) subset of HuggingFaceVLA/libero — 379 episodes across all 10 tasks, not a toy subset.
Downloads · 30 days
15
19% of all-time downloads
All-time downloads
81
Public
Parameters
2.4B
13.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors7 GB · 100%
How the weights are stored.
BF161.3B · 56%
From the Hugging Face model README
A real fine-tune of NVIDIA GR00T-N1.5-3B via LeRobot's
groot policy wrapper on the full LIBERO-Long (libero_10) subset of
HuggingFaceVLA/libero — 379 episodes across all 10 tasks, not a toy subset.
nvidia/GR00T-N1.5-3BHuggingFaceVLA/libero, episodes 0–378 (the 379 episodes belonging
to the 10 libero_10 tasks; the dataset's other 1,314 episodes, from
libero_spatial/libero_object/libero_goal/libero_90, were excluded).transformers/accelerate
incompatibility (ValueError: too many values to unpack (expected 0)
inside PyTorch module iteration, hit intermittently at various points) and
a dataset-caching bug where resuming via --config_path sometimes loaded
a stale/wrongly-sized cached copy of the dataset, causing a dataloader
IndexError. The second one was fixed by resuming with a fresh
HF_HOME cache directory. Final loss was ~0.07–0.11, essentially flat
from step ~15,000 onward.lora_rank=8 was passed but did not reduce the trainable parameter count in
this LeRobot integration (same caveat as our earlier 2-episode stress-test
checkpoint) — the diffusion head was fully fine-tuned, not LoRA-adapted.Our previous release (zhoumiaosen/groot-n1p5-lora-libero-2ep-stresstest)
was trained on only 2 episodes as a deliberate GPU/software stress test and
explicitly not a real policy. This checkpoint is a genuine full fine-tune
across the entire LIBERO-Long task suite, evaluated and compared against a
zero-shot GR00T-N1.5-3B baseline on the same tasks below.
Evaluated on the full libero_10 (LIBERO-Long) suite, 5 episodes per task,
with correctly calibrated action normalization (real min/max statistics from
HuggingFaceVLA/libero's dataset metadata, not identity normalization — see
note below on why that calibration step matters). Eval episodes use
different seeds than training, so this reflects generalization, not
memorization of the training set.
Overall: 6/50 = 12.0% success (vs. 0/50 = 0% for zero-shot GR00T-N1.5-3B, identically calibrated, no fine-tuning)
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 | 2/5 | 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 | 2/5 | 0/5 |
| 6 | put the white mug on the plate and put the chocolate pudding to the right of the plate | 1/5 | 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 | 1/5 | 0/5 |
Tasks 2, 3, 4, 5, 6, 9 are the 6-task subset used elsewhere in our project (the rest have only a single subgoal and aren't part of that study); on that subset this checkpoint scores 4/30 = 13.3%, vs. 0/30 zero-shot.
Absolute success rates are still low, consistent with GR00T needing substantial data/training to do well on long-horizon LIBERO tasks with only 379 training episodes. The meaningful result here is the direction: fine-tuning took this checkpoint from 0% to a real, non-trivial, non-memorized success rate on held-out-seed episodes, confirming the training → checkpoint → eval pipeline works end-to-end, and completing the full 30k-step recipe improved on the 25k-step interim checkpoint (8.0% → 12.0% overall).
An initial zero-shot attempt (default normalization, no dataset stats)
produced 0/50 with suspicious raw actions — all-positive, narrow-banded,
almost no variance, inconsistent with LIBERO's expected [-1, 1] action
range. Root cause: make_pre_post_processors needs an explicit
dataset_stats argument (real per-dimension min/max) to un-normalize GR00T's
output into the target robot's action space; without it, it silently
defaults to identity normalization. Re-running with real stats from
LeRobotDatasetMetadata("HuggingFaceVLA/libero").stats produced correctly-
scaled actions (confirmed by direct inspection) and — still 0/50. That
second number is the one reported above as the trustworthy zero-shot
baseline.