AI Keeps Getting Smarter—And Now Scientists Want to Know If It Can Build Robots

A research team led by Na Li and Bo Dai has developed a new open-source benchmark, RLE-Bench, to evaluate whether AI can handle the complex challenges of robotics engineering. The benchmark assesses if AI agents can design and implement robotic systems, including software, hardware, and control algorithms. It includes 48 tasks across four areas, such as sensor interaction, motor control, and physical constraints. The benchmark aims to create a common measuring stick for comparing AI systems in robotics.

RLE-Bench evaluates AI's ability to handle real-world engineering constraints, such as weight, torque, and balance. The benchmark is not just about learning control policies but also about designing hardware and interfaces. Li emphasizes that the current benchmark is only the beginning, and future versions will focus on broader engineering tasks. The benchmark is used in simulation environments where AI agents can refine their designs through trial and error.

China is advancing in robotics and AI, hosting the second World Humanoid Robot Games in 2026. The games featured over 2,000 robots from 666 teams across 16 countries, showcasing real-world applications like household tasks and emergency rescue. RLE-Bench and such competitions evaluate different aspects of robotics, making direct comparisons difficult. RLE-Bench focuses on AI's engineering capabilities, while physical competitions test real-world performance. Together, these approaches highlight the growing challenge of creating AI systems that can reason about physical machines and ensure reliable performance outside simulated environments.

AI-generated summary of the original article — see the source for the full story.