Learning to Design Soft Hands using Reward Models