FIRe-VLA-OpenVLA-PegInsert-Sim

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

• Framework: OpenVLA
• Task: Peg insertion (peg_insert)
• Domain: Sim
• Checkpoint: inference-ready (20000 steps)

Usage

# Requires the OpenVLA-OFT codebase: https://github.com/moojink/openvla-oft
from transformers import AutoModelForVision2Seq, AutoProcessor

repo = "bhe1004/FIRe-VLA-OpenVLA-PegInsert-Sim"
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForVision2Seq.from_pretrained(repo, trust_remote_code=True)
# OFT extras bundled in the repo: action_head, proprio_projector,
# lora_adapter, dataset_statistics.json

Related

• Training data: bhe1004/FIRe-VLA-OpenVLA-PegInsert-Sim