fatigue-injection-openpi

fatigue-injection-openpi

π0.5 (pi05) checkpoints fine-tuned from gs://openpi-assets/checkpoints/pi05_base/params with openpi on Franka Panda stack-cube rollout datasets (LeRobot v2.1 format), for studying fatigue injection.

Runs

Each run directory contains the final checkpoint (step 19999): params/ (model weights, orbax format) and assets/<asset_id>/norm_stats.json (normalization statistics).

Task prompts: the cmd_x0005 and cmd_x001 datasets use "stack the red cube yellow cube and green cube in order on the tray"; cmd_x0015, x001, and x002 use "stack the red cube and yellow cube in order on the tray".

Note: run directories are named seedN (dataset is identified by the parent directory) because the Hugging Face API rejects paths containing the sequence /x00 (e.g. x001_seed0/), which it treats as an escaped null byte.

Training setup

• Model: Pi0Config(pi05=True, action_dim=32, action_horizon=16)
• Init: pi05_base params
• Steps: 20,000 (checkpoint at final step 19999)
• Batch size: 32, single A100-80GB per run
• LR: cosine decay, warmup 1k, peak 2.5e-5, decay to 2.5e-6 over 30k
• Data: 3 cameras (2 exterior + 1 wrist, 480x640), 14-dim state (cartesian pose 6 + gripper used), 7-dim delta-EEF actions, prompt from task
• Norm stats computed per dataset (included under each run's assets/)

Usage

Download a run directory and load it with openpi's create_trained_policy, e.g.:

from openpi.policies import policy_config
from openpi.training import config as train_config

policy = policy_config.create_trained_policy(
    train_config.get_config("pi05_fatigue_x001"),
    "/path/to/pi05_fatigue_x001/seed0",  # dir containing params/ and assets/
)