oucliuyang/GAWM-S-LIBERO

GAWM-S for LIBERO

GAWM (Geometry-Aware World Model) is a compact multi-task world-action model. A lightweight predictor forecasts the patch features of a frozen DINOv3 ViT-S/16 encoder at two future horizons, and an action decoder regresses action chunks from the current and forecast tokens. GAWM uses neither a vision-language model nor a video generator.

This repository holds the LIBERO model reported in the paper.

Results

One policy for the four suites; all 50 initial states of each of the 40 tasks (2,000 episodes), seed 7.

Spatial	Object	Goal	Long	Average
96.2	99.8	95.0	91.2	95.55 (1,911 / 2,000; 95% interval [94.6, 96.4])

Model

Parameters	38.1M in total: DINOv3 ViT-S/16 encoder 21.6M (frozen, not in this repository) and adapter, predictor, task code and action decoder 16.5M (this checkpoint)
Inputs	third-person and wrist cameras at 256×256, 8-D state, and the task instruction, which selects its task code stored in the checkpoint
Output	8 steps of 6-D end-effector deltas and a gripper command
Latency	7.1 ms per action-chunk query (median of 200; batch size 1, bf16, NVIDIA RTX PRO 5000; from preprocessed inputs to the action chunk)

Files

File	Content
checkpoints/GAWM-S-LIBERO.safetensors	weights
config.yaml	model configuration
dataset_statistics.json	action and state normalization
SHA256SUMS	checksums of the files above

Usage

The model runs with the GAWM code at github.com/gaoxiang12/gawm-public; set up its environment as described in the README, then from the repository root:

hf download oucliuyang/GAWM-S-LIBERO --local-dir playground/Checkpoints/GAWM-S-LIBERO
(cd playground/Checkpoints/GAWM-S-LIBERO && sha256sum -c SHA256SUMS)

DINOv3 ViT-S/16 is downloaded to playground/Pretrained on first use. Loading logs a warning that the keys backbone.encoder.* and world_model.feature_decoder.* are missing. This is expected: the encoder weights come from the official DINOv3 release, and the feature decoder is not used at inference.

Closed-loop evaluation with the protocol above:

python scripts/run_libero_checkpoint_eval.py \
  --checkpoint playground/Checkpoints/GAWM-S-LIBERO/checkpoints/GAWM-S-LIBERO.safetensors \
  --output playground/Evaluations/gawm_s_libero_50ep \
  --gpus 0,1,2,3,4,5,6,7 --trials 50 --seed 7

License

MIT. DINOv3 is not included; it is distributed by Meta under the DINOv3 License.

Citation

The paper will be linked here once it is public.