GCA-Bench reveals robot grasping models fail 30% of complex tasks

Robotic grasping has long been a fundamental challenge for real-world applications, but most existing benchmarks only test isolated visual grasp pose detection—ignoring the multi-step reasoning and semantic understanding needed for complex tasks. To address this gap, a team of researchers from multiple institutions (including the University of Liverpool, Auburn University, and others) developed GCA-Bench, a new benchmark that introduces "grasping with complex action" scenarios. These scenarios require both scene-level reasoning (e.g., understanding object relationships) and semantic constraints (e.g., grasping a tool by the handle, not the blade).

GCA-Bench evaluates a diverse set of baselines, from traditional grasp detection pipelines to end-to-end learning methods and recent large foundation models. The results are sobering: even the best models achieve success rates below 70% on these complex grasping tasks, compared to near-perfect performance on simpler benchmarks. The paper proposes new evaluation metrics, analyzes critical failure modes (e.g., reasoning errors, misinterpreting object semantics), and offers insights to guide the development of more robust and generalizable grasping strategies. This work underscores that current AI systems still lack the integrated reasoning required for truly autonomous manipulation in unstructured environments.