fine-tuning-openvla-oft

Launch vla-scripts/finetune.py via torchrun on 8 GPUs with use_l1_regression=True, lora_rank=32, learning_rate=5e-4, and an RLDS dataset like libero_spatial_no_noops. Evaluate 50k, 100k, and 150k checkpoints and keep the best by task success rate.

OFT disables FiLM conditioning and uses 2 camera images, targeting LIBERO simulation. OFT+ enables FiLM for stronger language grounding and uses 3 image streams, targeting ALOHA real-world bimanual tasks.

Use OpenVLA-OFT when adapting the openvla-7b backbone with LoRA and continuous L1 or diffusion action heads on LIBERO or ALOHA. Choose OpenPI-based workflows when you need pi0 or pi0.5 model architectures instead.

Performance drops occur because merged adapter artifacts mismatch across hardware and runtime stacks. Re-merge the LoRA weights on the target machine with merge_lora_weights_and_save.py, then re-run evaluation with identical runtime flags.

Low success usually comes from train-eval config mismatches, most commonly missing center_crop=True when trained with image_aug=True. Also verify num_images_in_input, use_film, lora_rank, and unnorm_key parity between training and evaluation.

LIBERO evaluation needs about 16 GB VRAM on a single GPU, while ALOHA evaluation needs about 18 GB. Fine-tuning uses 8x A100 GPUs at roughly 27 GB per GPU for LIBERO and 35 GB per GPU for ALOHA OFT+.