jermo1120/b1k-spraying-for-bugs-pi05

π0.5 LoRA fine-tune for BEHAVIOR Challenge task 38: spraying_for_bugs

Pick up the pesticide atomizer in the garden and spray insectifuge to fully cover both potted plants in the garden.

Status: training in progress. checkpoint/ always holds the most recent saved step (see LATEST_STEP). Older checkpoints are replaced. Not yet evaluated in OmniGibson.

Recipe

Base model	π0.5 via the OpenPI behavior fork (base commit in code/openpi_base_commit.txt)
Initialization	Comet pi05-b1kpt50-cs32: 2nd place, 2025 BEHAVIOR Challenge; pretrained on all 50 2025 tasks
Data	behavior-1k/2026-challenge-demos chunk-038: 200 demos, 1,295,895 frames, RGB only (head + 2 wrist cams)
Fine-tuning	LoRA (gemma_2b_lora + gemma_300m_lora), batch 32, 50k steps, cosine LR 5e-5 → 5e-6 with 1k warmup, action horizon 32
Normalization	Comet's quantile norm stats (shipped under checkpoint/assets/)
State/action layout	Absolute actions (no delta transform); proprio ordered base, trunk, L arm, R arm, L grip, R grip (b1k/R1Pro_comet) to match Comet
Hardware	1× NVIDIA L40S 46 GB, ~4.2 s/step

Reproduce

git clone -b behavior https://github.com/wensi-ai/openpi.git && cd openpi
git checkout $(cat code/openpi_base_commit.txt) && git apply code/openpi_behavior_branch.patch
GIT_LFS_SKIP_SMUDGE=1 uv sync

# Download chunk-038 (RGB) of 2026-challenge-demos and the Comet checkpoint, then build the
# re-indexed single-task dataset (LeRobot can't load one chunk of the full dataset directly):
python code/make_task_dataset.py --src <2026-challenge-demos> --dst /workspace/b1k/spraying_for_bugs

HF_HUB_OFFLINE=1 XLA_PYTHON_CLIENT_MEM_FRACTION=0.92 uv run scripts/b1k/train_b1k.py \
    pi05_b1k_spraying_for_bugs_comet_lora --exp_name=<name> --overwrite

Paths inside the config (/workspace/...) need to match your machine.

Files

• checkpoint/: latest OpenPI/Orbax checkpoint (params/, train_state/, assets/ norm stats)
• code/openpi_behavior_branch.patch: config, robot layout and task prompt changes against the fork
• code/make_task_dataset.py: turns one downloaded chunk into a standalone LeRobot v3 dataset
• code/hf_uploader.sh: pushes each new checkpoint here during training