hermanprawiro/ebim_task2_pi05_fft_fixpos_200

pi05_ebim_task2_fixpos_200_full_finetune (step 29999)

openpi training checkpoint for EBiM task2 simulation phase. Trained on 200 episodes of simulation data. The robot base does not move and no object randomization.

Train config	pi05_ebim_task2_fixpos_200_full_finetune
Step	29999
Norm stats	assets/ebim/task2_fixpos_200/norm_stats.json
Framework	openpi (JAX)

Layout

• params/ - JAX/orbax checkpoint of the inference parameters. Loadable only through openpi (openpi.models.model.restore_params), not through transformers or safetensors.
• assets/ebim/task2_fixpos_200/norm_stats.json - normalization statistics used at training time.
• _CHECKPOINT_METADATA - orbax bookkeeping.
• docs/ebim_task2_policy.py - the EBIM task2 input/output transforms (drop-in openpi module).
• docs/train_config.py.txt - the data config class and the pi05_ebim_task2_fixpos_200_full_finetune train config entry, to add to openpi's config.py.
• docs/source.json - config name, step and asset id.

Serving

Starting from a fresh clone of openpi:

1. Copy docs/ebim_task2_policy.py to src/openpi/policies/ebim_task2_policy.py.
2. Add the contents of docs/train_config.py.txt to src/openpi/training/config.py: the import line next to the other imports, the data config class next to the other DataConfigFactory subclasses, and the TrainConfig(...) entry inside the _CONFIGS list.
3. Download this repo into the checkout and start the policy server:

uv run scripts/serve_policy.py policy:checkpoint \
    --policy.config=pi05_ebim_task2_fixpos_200_full_finetune --policy.dir=checkpoints/hermanprawiro/ebim_task2_pi05_fft_fixpos_200