You can deploy this model and start earning money today!

EmptyDrum/
sharpen-pencil-pi05-v2

johannesmichalke/sharpen-pencil

lerobot/pi05_base

adapter

apache-2.0

Views

No views yet

sharpen-pencil-pi05-v2

LoRA finetune of lerobot/pi05_base on johannesmichalke/sharpen-pencil — a YAM single-arm 7-DOF dataset of "move a pencil from the box into the sharpener."

Repo layout

Each subfolder is a complete checkpoint at the corresponding training step. Pick one when loading.

subfolder	step	training loss	epoch
`010000/`	10 000	0.021	1.25
`020000/`	20 000	0.015	2.50
`030000/`	30 000	0.012	3.74

All three are LoRA adapters of the same shape (~369 MB each); pick 030000/ for best fit.

Quick load

Requires Python 3.12, lerobot[pi,dataset] (≥ 0.5.1), and peft (≥ 0.18). No HF auth needed — public repo.

python

from huggingface_hub import snapshot_download
from lerobot.policies import make_policy

ckpt = snapshot_download(
    "EmptyDrum/sharpen-pencil-pi05-v2",
    allow_patterns="030000/*",
)
policy = make_policy(pretrained_policy_name_or_path=f"{ckpt}/030000")
policy.eval()
# policy.predict_action(observation) -> (10, 7) action chunk

Base model lerobot/pi05_base (~6 GB) is fetched lazily on first construction.

What the model expects

Cameras (3 keys, only 2 carry information):
- observation.images.base_0_rgb — top-down view (trained on Logitech, cam_high in dataset)
- observation.images.left_wrist_0_rgb — wrist view (trained on RealSense D405, cam_wrist in dataset)
- observation.images.right_wrist_0_rgb — always masked during training. Pass a zero image with image_mask=False, or set policy.config.empty_cameras=1.
State: observation.state = float32[7] = [joint0…joint5, gripper]. Joints in radians, gripper ∈ [0, 1]. State is padded to 32 dims internally by the policy — feed 7 raw values.
Action: model outputs a chunk of 10 actions, each padded internally; the postprocessor unpads back to 7 dims in the same units.
Prompt: the training task string is "move a pencil from the box into the sharpener". It comes from the dataset's meta/tasks.parquet automatically when using the lerobot data loader, but for live inference pass it explicitly in the observation dict.

Training details


Base model	`lerobot/pi05_base`
Dataset	`johannesmichalke/sharpen-pencil` (LeRobot v3.0, 77 episodes, 64 107 frames, 30 fps)
Method	LoRA via PEFT, `target_modules=all-linear`, `r=32`, `lora_alpha=32`
Trainable / total params	92 M / 4.2 B
Batch size	8
Steps	30 000 (≈ 3.74 epochs)
LR schedule	Cosine decay with 1 k warmup, peak `2.5e-5`, decay to `2.5e-6`
Optimizer	AdamW, grad-clip 1.0
Precision	bf16 with `torch.compile=max-autotune` + gradient checkpointing
Action chunk	10
`max_token_len`	200 (pi05 default)
Hardware	1× H100 80 GB
Wall-clock	10 h 34 min

Caveats

Not yet evaluated on a real YAM. Loss-curve numbers above are training-only; no closed-loop / dataset-holdout eval has been run.
The right-wrist camera slot is preserved in the model's input schema purely for compatibility with the pi05_base API; it carries no information from this dataset.
LoRA r=32 means the adapter fits any pi05_base-compatible model, but the deltas are tied to this particular task + robot + camera layout. Don't expect it to generalize.

Reproducing

Full training config lives at HalfEmptyDrum/Server-YAM:

runs/sharpen_pencil_v2.env — every hyperparameter for this run
scripts/train.sh — generic lerobot-train launcher
README in the same repo has the full setup recipe (one-time uv + lerobot install, then ./scripts/train.sh sharpen_pencil_v2).

License

Apache 2.0 — same as the upstream lerobot/pi05_base checkpoint.