DuoBench: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World

DuoBench is a newly introduced benchmarking framework designed for evaluating bimanual manipulation policies on the FR3 Duo robot platform. It features eleven tasks across four coordination categories, with implementations in both simulation and real-world settings, supported by a stage-based evaluation scheme for detailed failure analysis. The framework highlights the challenges faced by current dual-arm imitation-learning and vision-language-action policies, particularly in early interactions and real-world transfer, providing a valuable resource for practitioners focused on improving dual-arm robotic systems.