Hugging Face's logo

LeWAM
/
lewam-drawer

Model card Files
xet
 Community

LeWAM Drawer Cleanup

Goal-reaching checkpoint for DexMimicGen TwoArmDrawerCleanup in the LeWAM v1 format.

Files 

lewam_best.pt
lewam_config.json

Download 

export STABLEWM_HOME="${STABLEWM_HOME:-$HOME/.stable_worldmodel}"
hf download LeWAM/lewam-drawer --local-dir "$STABLEWM_HOME/checkpoints/lewam-drawer"
hf download LeWAM/lewam-drawer drawer.h5.zst --repo-type dataset --local-dir "$STABLEWM_HOME/datasets"
zstd -d --long=27 "$STABLEWM_HOME/datasets/drawer.h5.zst"

The single-view dataset contains model_xml for all 1,026 episodes. The v1 environment uses it to restore each trajectory's payload geometry. The three-view dataset exposes the same column through drawer_multiview.h5.

Evaluation 

python scripts/eval_lewam.py --config-name drawer policy=lewam-drawer seed=42 \
    eval.mode=lewam_plan eval.plan_mode=grad \
    eval.grad_all_k=true eval.grad_tr=0.01 eval.exec_actions=5

Goal offset: 50 steps. Action budget: 100 steps. Evaluation: 50 episodes per seed. Repeat with seed=0 and seed=1 for the three-seed results.

SHA256 of lewam_best.pt: c91ddb9ec2cafa95e4166fdbec4093c7b4ff67bede8ee36bd12ee2598494f4f6.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including LeWAM/lewam-drawer