FM-based reward design for long-horizon robotics

Develop foundation-model-based reward design methods that are effective for long-horizon robotic tasks, providing reinforcement signals suitable for reinforcement learning in these settings.

Background

The paper surveys prior work that leverages LLMs and vision-LLMs to generate or shape reward functions for robotics, noting successes primarily in short-horizon or repetitive skills.

It explicitly states that applying foundation-model-based reward design to long-horizon tasks remains unresolved and highlights limitations of alternatives that rely on sparse success signals or trained reward models with limited scalability and task-specific engineering.