DuoBench: A Reproducible Benchmark for Bimanual Manipulation in Simulation and the Real World
Existing benchmarks struggle to effectively evaluate the control complexity and failure modes inherent in dual-arm collaborative manipulation. To address this gap, this work proposes DuoBench—the first reproducible evaluation framework specifically designed for dual-arm tasks—encompassing 11 distinct tasks across four categories of collaboration and supporting deployment in both simulation and real-world environments. DuoBench introduces staged fine-grained assessment, cross-modal policy evaluation spanning vision, language, and action, and a human teleoperation dataset, with task consistency ensured through the FR3 Duo platform and 3D-printed assets. Experimental results demonstrate that current policies exhibit significant limitations in early-stage interaction, concurrent bimanual execution, and sim-to-real transfer, and that DuoBench can effectively uncover and semantically interpret these failure modes.