yeeeiii111/wuji-writing-pi0

Wuji Writing Pi0 (action_dim=54)

Pi0 finetune on the Wuji 毛笔 (Chinese calligraphy brush) writing dataset. Dual ARX-5 + dual dexterous hands, 54-dim state/action, 3 cameras (stereo_left as the base + left/right wrist).

• Base ckpt: pi0_base_pytorch with the 3 action_dim-bound projection layers (state_proj, action_in_proj, action_out_proj, with their biases) reinitialised to fit 54-D — see scripts/convert_base_ckpt_action_dim.py
• Dataset: yeeeiii111/wuji-writing (teleop subset, 487 ep → 439 train / 48 val with seed=42, val_ratio=0.1, byte-for-byte aligned with VAM's split)
• Training repo: https://github.com/Qingxin-Wang/openpi-writing (branch remote/pi05_writing — single branch carries pi05 + pi0 baselines)
• Schedule: 16-GPU run, global batch 32, lr 5e-5 with 500-step warmup

Available checkpoints

Each subdirectory is a self-contained inference-ready ckpt (model.safetensors + metadata.pt + assets/).

Path	Training step	Approx. epochs
25k/	25,000	~2
50k/	50,000	~4 (pending upload when training reaches step)

optimizer.pt is intentionally not uploaded (~12.5 GB). These ckpts are for inference / serving; to resume training, retrain from pi0_base_pytorch_a54 or contact the authors.

Use

# Download one of the steps -- here 25k
huggingface-cli download yeeeiii111/wuji-writing-pi0 \
  --include "25k/*" --local-dir ./pi0_writing
# (the contents land at ./pi0_writing/25k/{model.safetensors, metadata.pt, assets/...})

# Serve from the openpi repo (Qingxin-Wang/openpi-writing@remote/pi05_writing):
.venv/bin/python scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi0_writing \
  --policy.dir=./pi0_writing/25k \
  --port=8000

Client side: see docs/writing_baseline.md.