Reward Design for Physical Reasoning in Vision-Language Models

Develop a lightweight reward-shaping library specifically for VLMs focused on physics-based tasks. The library should allow users to easily toggle different reward structures for fine-tuning agents on common 3D physics simulators like MuJoCo or IsaacGym.

Suggested repo: physreward

"Make your VLMs actually understand gravity and collision."

Estimated effort: 45h