A minimal reimplementation of VLA-0 using TRL's SFTTrainer that achieves ~90% performance on LIBERO...