PartDexGrasp — checkpoints

Model weights from the PartDexGrasp repo. All of them are model weights only (optimizer/scheduler removed, EMA weights in place of the raw ones): sufficient for evaluation / prediction generation, not for resuming training.

`dexvlg_v2_masked37_best_mingt.pt`

DexVLG — Qwen3-0.6B trunk + π0-style expert on a Part-VAE latent, masked-latent arm (live 37/112), step 28,500, best minGT 58.3 mm. The checkpoint stores cfg (incl. live_mask), so the loader reconstructs the masked-latent setup automatically.

huggingface-cli download leesuyun/partdexgrasp-dexvlg \
    dexvlg_v2_masked37_best_mingt.pt --local-dir checkpoints/dexvlg
python eval/gen_qwen_predictions.py \
    --ckpt checkpoints/dexvlg/dexvlg_v2_masked37_best_mingt.pt \
    --out preds/predictions.json

`kppersp_smol500m_vary16_ft20_best.pt`

kppersp — SmolVLM-500M trunk + π0-style expert that flow-matches 2-D hand keypoints in an L515 pinhole camera frame; kp2pose lifts them to a 31-D ShadowHand pose. Variable-view arm: one model covering 3-6 input views, fine-tuned for 20 epochs from the fixed 6-view arm with --vary_view 1,2,3,4,5,6 (the 6-view-only model collapses when views are dropped at test time).

epoch 18 of 20, selected by test/mean_3d_mm = 62.08
2.18 GB; holds model / args / ep / mean_3d_mm / cam_dim
the --kp_std statistics ride along as the kp_mean / kp_scale buffers inside the state dict, so no separate kp_stats.npz is needed at inference
the flags the loader must match are listed in checkpoints/kppersp/README.md

huggingface-cli download leesuyun/partdexgrasp-dexvlg \
    kppersp_smol500m_vary16_ft20_best.pt --local-dir checkpoints/kppersp

`dexgys_final/`

Precomputed caches (3-view hand keypoints, Janus goal tokens, SigLIP object features) used by the arms above; see the repo's preprocessing/ for how they are built.

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