fine-tuning-openvla-oft

Fine-tune OpenVLA-OFT for robot action generation by applying LoRA adaptation to the VLA backbone with optional FiLM conditioning and continuous action heads. It supports training custom action heads using L1 or diffusion approaches across LIBERO and ALOHA setups.

Yes, OpenVLA-OFT supports continuous action heads for ALOHA deployment using L1 or diffusion models. It enables server-client inference workflows and allows LoRA merging to resolve cross-GPU normalization issues during deployment.

Debug cross-GPU normalization issues during OpenVLA-OFT evaluation by matching training and inference flags and ensuring action head, FiLM, image stream, and LoRA rank parity. This maintains inference invariants across different hardware configurations.

Yes, you can use FiLM conditioning with LoRA-based fine-tuning on the VLA backbone, known as OFT+. This approach allows multi-image input and supports continuous action head training for both LIBERO and ALOHA pipelines.

OpenVLA-OFT inference fails after merging LoRA weights across different GPUs due to broken invariants. Ensure action head, FiLM, image stream, and LoRA rank parity are maintained, and verify training and inference flags match exactly.