FIRe-VLA-PI05-PegInsert-Real

A PI05 vision-language-action policy fine-tuned for the Peg insertion contact-rich robotic assembly task (Real domain), part of the FIRe model suite.

• Framework: PI05
• Task: Peg insertion (peg_insert)
• Domain: Real
• Checkpoint: inference-ready (2999 steps)

Usage

# Requires openpi: https://github.com/Physical-Intelligence/openpi
huggingface-cli download bhe1004/FIRe-VLA-PI05-PegInsert-Real --local-dir ./FIRe-VLA-PI05-PegInsert-Real

# Point the openpi trained-policy loader at the downloaded checkpoint dir.
# The repo holds params/ + assets/ (norm stats); train_state/ was dropped
# since it is only needed to resume training.
from openpi.policies import policy_config
policy = policy_config.create_trained_policy(config, "./FIRe-VLA-PI05-PegInsert-Real")

Related

• Training data: bhe1004/FIRe-VLA-PI05-PegInsert-Real