XLANG Lab launches RoboWild benchmark for robot-use agents
The two-year gap between OSWorld and RoboWild benchmarks shows a pattern: each new frontier benchmark initially crushes frontier models, then becomes the roadmap labs optimize against. Robotics researchers and embodied-AI teams now have an open, reproducible evaluation suite spanning four robot embodiments, downloadable from Hugging Face and GitHub. Teams building physical-world agents can use the 341-task suite to measure generalization gaps before deploying in homes, stores or workshops.