## Motivation Full fine-tuning of pi0 models is expensive. LoRA would enable efficient task-specific adaptation with minimal VRAM. ## Proposal Add a LoRAPolicy wrapper that applies LoRA adapters to attention layers while keeping the action head trainable.
Motivation
Full fine-tuning of pi0 models is expensive. LoRA would enable efficient task-specific adaptation with minimal VRAM.
Proposal
Add a LoRAPolicy wrapper that applies LoRA adapters to attention layers while keeping the action head trainable.