siruku6/pi05_all111

π0.5 — all-task combined fine-tune (2,000 steps)

A full-parameter fine-tune continuing from siruku6/pi05_full_runpod (rp3000), trained on a combined LIBERO + LIBERO-plus dataset resampled to 20 Hz so the policy sees the full task set in a single run. The scene camera is passed as an empty placeholder (empty_cameras=1); no camera dropout is applied.

No evaluation results are reported here. This run was trained offline (eval_steps=0) and was not scored on held-out episodes, so nothing in this card is a performance claim — it is a training artifact only.

Lineage

Stage	Weights	Steps	Note
Base	lerobot/pi05_libero_base @ a217bfd3b14673cf2ce597e69997ab21866438dd	—	π0.5, 4.14 B params
+1	siruku6/pi05_stage1_24000	24,000	action expert only
+2	siruku6/pi05_full_runpod (= rp3000)	3,000	full-parameter fine-tune
+3 (this repo)	000500 … 002000	2,000	full-parameter, combined LIBERO + LIBERO-plus @ 20 Hz

Contents

Four checkpoints, saved every 500 steps. Each directory is the flattened content of a LeRobot pretrained_model/: config.json, model.safetensors (9,354,050,752 bytes), train_config.json. logs/pi05_all111.log is the full training log.

000500/  001000/  001500/  002000/  logs/

Training setup

Objective / policy	π0.5 (pi05), flow-matching action expert, action chunk 50, n_action_steps 10
Trainable	all 4,143,404,816 parameters (freeze_vision_encoder=false, train_expert_only=false)
Batch / LR	64 / 5e-6 peak → 2.5e-6, cosine decay with warmup (warmup 1000 auto-scaled to 66 for 2,000 steps)
Precision	bfloat16, gradient checkpointing on
Seed	42
Cameras	wrist + front; scene camera passed as an empty placeholder (empty_cameras=1), no dropout
Data	combined LIBERO + LIBERO-plus resampled to 20 Hz — 3,028,708 frames, 19,533 episodes (local assembly, not published on the Hub)
Hardware / time	1× NVIDIA RTX PRO 6000 Blackwell (95 GB), 3 h 48 m 43 s, 6.81 s/step, 49.4 GB VRAM
Framework	LeRobot (π0.5 is a port of the OpenPI implementation), seed 42

Use

hf download siruku6/pi05_all111 --include '002000/*' --local-dir ./all111

Then point LeRobot at the local directory:

lerobot-train --policy.type=pi05 --policy.pretrained_path=./all111/002000 ...

Note that the pretrained_model/ level is flattened away in this repo, so lerobot-train --resume against the Hub path will not find train_config.json where it expects it. Downloading first and passing a local --policy.pretrained_path works.

Intended use and limitations

Research artifact. Trained only on LIBERO / LIBERO-plus simulation data with a simulated Franka Panda — there is no real-robot validation, no held-out evaluation, and nothing here should be run on physical hardware without your own safety review. Performance outside the LIBERO task and camera setup is unknown.

License

These weights are a Model Derivative of Gemma (via PaliGemma inside π0.5) and are released under the Gemma Terms of Use. Use is also subject to the Gemma Prohibited Use Policy. See NOTICE for the third-party attributions that come with the base model, the dataset and the training code.