DrEureka is a novel algorithm that leverages large language models (LLMs) to automate and accelerate sim-to-real transfer in robotics. It addresses two key challenges in sim-to-real design: reward function design and domain randomization parameter configuration. DrEureka uses an LLM to synthesize effective reward functions and domain randomization configurations, enabling efficient and safe real-world transfer of policies learned in simulation. The algorithm first generates reward functions using Eureka, a state-of-the-art LLM-based reward design method, and incorporates safety instructions to ensure safer behavior. It then constructs reward-aware physics priors by evaluating policies on perturbed simulations, which guide the LLM in generating domain randomization configurations. Finally, the LLM generates DR configurations based on these priors, allowing for effective real-world deployment of policies. DrEureka has been evaluated on quadruped locomotion and dexterous manipulation tasks, demonstrating its effectiveness in achieving competitive performance with human-designed configurations. On quadruped locomotion, DrEureka-trained policies outperform human-designed ones by 34% in forward velocity and 20% in distance traveled across various terrains. In dexterous manipulation, DrEureka's best policy performs nearly 300% more in-hand cube rotations than human-designed policies. DrEureka also successfully solves a novel task, the walking globe, where a quadruped balances and walks on a yoga ball for extended periods in both simulation and the real world. DrEureka's approach is validated through extensive experiments and ablation studies, showing that all components of the algorithm are essential for generating effective safety-regularized reward functions. The algorithm is capable of automating the sim-to-real transfer process without human intervention, making it a promising tool for accelerating robot learning and deployment in the real world. DrEureka's ability to handle complex tasks and its effectiveness in real-world environments highlight its potential as a versatile solution for sim-to-real transfer in robotics.DrEureka is a novel algorithm that leverages large language models (LLMs) to automate and accelerate sim-to-real transfer in robotics. It addresses two key challenges in sim-to-real design: reward function design and domain randomization parameter configuration. DrEureka uses an LLM to synthesize effective reward functions and domain randomization configurations, enabling efficient and safe real-world transfer of policies learned in simulation. The algorithm first generates reward functions using Eureka, a state-of-the-art LLM-based reward design method, and incorporates safety instructions to ensure safer behavior. It then constructs reward-aware physics priors by evaluating policies on perturbed simulations, which guide the LLM in generating domain randomization configurations. Finally, the LLM generates DR configurations based on these priors, allowing for effective real-world deployment of policies. DrEureka has been evaluated on quadruped locomotion and dexterous manipulation tasks, demonstrating its effectiveness in achieving competitive performance with human-designed configurations. On quadruped locomotion, DrEureka-trained policies outperform human-designed ones by 34% in forward velocity and 20% in distance traveled across various terrains. In dexterous manipulation, DrEureka's best policy performs nearly 300% more in-hand cube rotations than human-designed policies. DrEureka also successfully solves a novel task, the walking globe, where a quadruped balances and walks on a yoga ball for extended periods in both simulation and the real world. DrEureka's approach is validated through extensive experiments and ablation studies, showing that all components of the algorithm are essential for generating effective safety-regularized reward functions. The algorithm is capable of automating the sim-to-real transfer process without human intervention, making it a promising tool for accelerating robot learning and deployment in the real world. DrEureka's ability to handle complex tasks and its effectiveness in real-world environments highlight its potential as a versatile solution for sim-to-real transfer in robotics.