Reward as An Agent for Embodied World Models

This work introduces a novel framework called Reward as an Agent for enhancing reinforcement learning (RL) in embodied world models. It integrates an agentic reward system that provides reliable verification of behaviors to prevent reward hacking, alongside Dynamic-Aware Rollout Diversification (DynDiff-GRPO) to enhance exploration through diversified action-space trajectories. This approach demonstrates significant accuracy improvements in RL applications, highlighting the potential for broader exploration strategies grounded in robust verification mechanisms, which is critical for practitioners aiming to develop more effective and resilient AI systems.