calvin-rdvla-lora-epoch0

RD-VLA (OpenVLA-7B recurrent LoRA, k_iters=8) fine-tuned on CALVIN task_D_D — first post-epoch-0 checkpoint (step16151).

Fine-tuned with LoRA on CALVIN task_D_D for the paper:

"Reasoning as a Double-Edged Sword: Vulnerability and Defense in VLA Pipelines" NeurIPS 2026 — University of Melbourne Physical AI Group

Code and probe results: https://github.com/uom-physical-ai/calvin-vla-adversarial

Checkpoint details

Field	Value
Base model	openvla/openvla-7b
Fine-tune dataset	CALVIN task_D_D (512,077 episodes)
Step	16151
Epoch (training)	0
Adversarial probe step	16151

Adversarial probe results (Gaussian image noise, n=300, seed=42)

σ	mean_RAS	p95_RAS	mean_RS
0.05	0.0158	0.0810	0.0354
0.10	0.0272	0.1611	0.0629
0.20	0.0673	0.2969	0.1227

RAS = Relative Action Shift (L2 deviation of predicted action, normalised). Lower = more robust.

Usage

from peft import PeftModel
from transformers import AutoModelForVision2Seq, AutoProcessor

base = AutoModelForVision2Seq.from_pretrained("openvla/openvla-7b")
model = PeftModel.from_pretrained(base, "uom-physical-ai/calvin-rdvla-lora-epoch0")
processor = AutoProcessor.from_pretrained("openvla/openvla-7b")

See the GitHub repo for the full fine-tuning and probe pipeline.