Tacoin fine-tuned GR00T checkpoint trained on the LIBERO libero long benchmark. The model follows the libero_gr00t data config (dual RGB streams + 8-DoF state) and predicts 16-step joint-space actions.

Training Snapshot

Base model: nvidia/GR00T-N1.5-3B
Checkpoint step: 7500 / 8000
Dataset: libero_long (10 tasks, 379 demos @ 10.0 FPS)
Run notes: long-horizon manipulation suite finetune

Evaluation

Offline reconstruction evaluated on 10 evenly spaced trajectories (160 steps each) with decord video backend and denoising_steps=4. Metrics are on unnormalized actions.

Metric	Value
Average MSE	0.03294
Median MSE	0.03433
Std MSE	0.01303
Max MSE	0.05386
Fraction ≤ 0.05	80.0%
Fraction ≤ 0.075	100.0%
Fraction ≤ 0.10	100.0%

Usage

python

from gr00t.experiment.data_config import load_data_config
from gr00t.model.policy import Gr00tPolicy

ckpt = 'Tacoin/GR00T-N1.5-3B-LIBERO-LONG'
data_config = load_data_config('libero_gr00t')
policy = Gr00tPolicy(
    model_path=ckpt,
    modality_config=data_config.modality_config(),
    modality_transform=data_config.transform(),
    embodiment_tag='new_embodiment',
    denoising_steps=4,
)

Feed a LeRobot-format observation dict into policy.get_action(...) to get a 16-step chunk.

Files

Path	Description
`config.json`	Transformer config for the action head.
`model-0000x-of-00002.safetensors`	Sharded weights.
`model.safetensors.index.json`	Weight shard index.
`experiment_cfg/metadata.json`	Dataset statistics for normalization.
`optimizer.pt`, `scheduler.pt`, `rng_state.pth`	Optimizer state for resuming.
`trainer_state.json`	Trainer snapshot.

License

Released under Apache-2.0; please cite NVIDIA Isaac GR00T and the LIBERO benchmark.