New preprint: SLowRL combines low-rank adaptation with a recovery policy for safer, more efficient fine-tuning of locomotion policies on hardware. [paper]