Tacoin fine-tuned GR00T checkpoint trained on the LIBERO libero spatial 8k benchmark. The policy consumes dual RGB views (video.image, video.wrist_image) plus an 8-D state and predicts 16 joint-space actions using the diffusion head.

Training Snapshot

Base model: nvidia/GR00T-N1.5-3B
Checkpoint step: 8000 / 8000
Dataset: libero_spatial (10 tasks, 432 demos @ 10.0 FPS)
Run notes: 8k-step continuation with headless evaluation fixes

Evaluation

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

Metric	Value
Average MSE	0.04587
Median MSE	0.04480
Std MSE	0.01975
Max MSE	0.08393
Fraction ≤ 0.05	70.0%
Fraction ≤ 0.075	90.0%
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-SPATIAL-8K'
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,
)

Pass a LeRobot observation dict to policy.get_action(...) to obtain the 16-step plan.

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 (loss curves, etc.).

License

Apache-2.0. Please credit NVIDIA Isaac GR00T and LIBERO when using this checkpoint.