IAAI233/UWM-LIBERO

UWM LIBERO checkpoints

This repository contains the two final UWM checkpoints selected after evaluation on LIBERO and LIBERO-Plus. The checkpoints are inference packages: model weights, sanitized experiment configuration, normalization statistics, trainer metadata, and aggregate evaluation results are included; optimizer states, schedulers, logs, videos, datasets, and text-embedding caches are not included.

The implementation is available in the UWM repository at source commit 28aae1ce746b87ddb36daa0f03ac8ec648d43a59. Both variants use Wan2.2-TI2V-5B as the pretrained backbone.

Included variants

Directory	Description	Training step
uwm-horizontal-21360	Selected horizontal two-camera UWM checkpoint	21,360
uwm-horizontal-no-video-action-21360	Ablation with video_action_conditioning: none	21,360

Each model-state file is stored at checkpoint-21360/pytorch_model/mp_rank_00_model_states.pt, matching the checkpoint layout expected by UWM. It is a DeepSpeed model-state payload whose module mapping contains the inference weights.

LIBERO results

Success rates are computed from 500 episodes per suite (2,000 episodes per variant).

Variant	Spatial	Object	Goal	Long	Overall
UWM horizontal	96.80	98.60	96.20	92.20	95.95
No video-action	96.00	98.20	95.20	87.00	94.10

LIBERO-Plus results

Success rates are shown by perturbation category. Overall is the exact episode-weighted aggregate across all 10,030 evaluated episodes, not an unweighted mean of the seven category percentages.

Variant	Background (1,076)	Camera (1,599)	Language (1,537)	Light (1,142)	Layout (1,525)	Robot (1,550)	Noise (1,601)	Overall (10,030)
UWM horizontal	55.95	18.39	77.29	78.11	67.48	45.23	63.02	56.98
No video-action	51.12	16.07	72.87	84.59	61.18	43.23	53.34	53.34

Exact successes, episode counts, suite aggregates, and full-precision rates are in each variant's eval/ directory. Artifact sizes and SHA-256 digests are recorded in manifest.json.

Download and load

Install the UWM code and its dependencies, then download this repository:

git clone https://github.com/Selen-Suyue/uwm.git
cd uwm
python -m pip install -e ".[train]"
hf download IAAI233/UWM-LIBERO --local-dir /absolute/path/to/UWM-LIBERO

For example, serve the selected horizontal checkpoint from the UWM checkout:

MODEL_DIR=/absolute/path/to/UWM-LIBERO/uwm-horizontal-21360
python -m deploy.libero.policy_server \
  --config "$MODEL_DIR/config.yaml" \
  --checkpoint "$MODEL_DIR/checkpoint-21360" \
  --device cuda:0 \
  --override \
  data.action_stats_path="$MODEL_DIR/libero_stats.json" \
  data.state_stats_path="$MODEL_DIR/libero_stats.json"

Use uwm-horizontal-no-video-action-21360 as MODEL_DIR to load the ablation. The configuration intentionally leaves training datasets and text-cache paths unset; they are not required for policy inference.

License

The packaged UWM material is released under Apache-2.0. The Wan2.2 base model and benchmark dependencies retain their own licenses and terms.