xhand cross-hand pi0.5 (no one-hot, arms Z / D) β€” fine-tuned deltas

pi0.5 fine-tuned on DexVerse Dexverse-GraspKettle-v0 with three pooled hands (sharpa, shadow, wuji) in the shared 41-D cross-hand action space [palm (or EE_common) position 3, rot6d 6, finger latent z 32], without a hand one-hot. Vision tower and action expert were trained; the PaliGemma language model was frozen.

These files are deltas, not full models. trainable.safetensors holds only the 647 trained tensors (844.9 M parameters). The pi0.5 base weights are not included. Get them from lerobot/pi05_base, under that model's own license. The deltas are fine-tuned from those weights, so their use may also fall under the base license.

Contents

folder arm action / state codec / f_task step
pooled3hand_nohot/Z/023670/ Z: palm pose (world) + E2* z 41 / 41 E2* / – 023670 (final, epoch 10)
pooled3hand_nohot/D/023670/ D: EE_common = palm @ inv(f_task(q_finger)) + E2* z 41 / 41 E2* / f_task P0a 023670 (final, epoch 10)

Each folder: trainable.safetensors, config.json (chunk_size 50, n_action_steps 8, max_state_dim/max_action_dim 41), train_config.json, pre/post-processor normalizers, export_meta.json, CODEC.txt. There was no snapshot selection: the final export is used, so best = last and it is uploaded once.

Results (final export, 50 episodes per hand, 3000-step cap)

arm sharpa shadow wuji total /150
Z 49 48 46 143
D 46 47 42 135

For comparison, the image DP with the same Z' / D' vectors scores 83 / 65 out of 150 (see xhand-crosshand-dp).

How to load (merge the delta onto the base)

export_meta.json β†’ merge: PI05Policy(config) <- base model.safetensors (stock key fix, RE-INIT keys dropped) <- trainable.safetensors.

  1. Build PI05Policy from this folder's config.json.
  2. Load model.safetensors of lerobot/pi05_base, applying LeRobot's key fix (policy._fix_pytorch_state_dict_keys), and skip action_in_proj / action_out_proj (re-initialized for the 41-D action space).
  3. Load trainable.safetensors on top (keys are prefixed model.). This also sets action_in_proj, action_out_proj, time_mlp_*.
  4. Optional check: export_meta.json β†’ frozen_sha16 lists sha256[:16] of three frozen weights (as float32 bytes) to confirm the right base.
from huggingface_hub import snapshot_download
d = snapshot_download("Dexverse-CSI/xhand-crosshand-pi05", allow_patterns=["pooled3hand_nohot/Z/023670/*"])
base = snapshot_download("lerobot/pi05_base")
# Steps 1-4 are implemented in the GitHub code, code/pi05_policy_server.py (functions build_policy, stream, load),
# which takes --ckpt <export folder> --base <pi05_base folder>; it also needs --port and --authkey; --check_only loads, runs one dummy chunk and exits.

export_meta.json β†’ base is the path used at training time; pass your own base path instead.

License: not yet specified (other); the pi0.5 base license applies to the base weights.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading

Model tree for Dexverse-CSI/xhand-crosshand-pi05

Finetuned
(899)
this model