The Sim-to-Real Gap of Foundation Model Agents: A Unified MDP Perspective

In three linesFoundation model agents suffer from sim-to-real gap. This paper formalizes the problem through MDP framework (observation, action, transition, reward) and proposes adapting classical robotics solutions (domain randomization). Example: multilingual tool calling reveals observation space gaps causing operationally invalid actions despite correct semantic intent.

Summary generated by Claude — human-verified