Evaluating progress in robotics is complex due to the interactive nature of the field. Unlike static domains like image classification, where datasets like ImageNet serve as benchmarks, robotics requires testing in dynamic environments. Small errors in action predictions can accumulate, leading to significant deviations in outcomes. As a result, researchers face two primary options for evaluation: using simulations or conducting real-world tests. Each method has its advantages and challenges. Simulations are improving in quality and diversity, with benchmarks like Libero and Calvin for manipulation tasks and Behavior 1k for mobile robotics. However, creating accurate simulations remains difficult, as they often lack the noise and unpredictability of real-world scenarios. Real-world evaluations are essential but logistically challenging. Unlike simulations, where environments can be reset easily, real-world tests require significant setup and cleanup. Inspired by the rapid progress in large language models, new evaluation methods like RoboArena are emerging. This community-run platform allows for cloud-based testing of robot policies, minimizing the number of costly evaluations needed. The upcoming Humanoid Everyday dataset also aims to facilitate real-world testing on humanoid robots, although details remain limited. Both approaches reflect a growing need for rigorous benchmarks that accurately reflect a robot's performance in practical applications.