ATOM-Bench: A Real-World Benchmark for Atomic Skills and Compositional Generalization in Manipulation Policies
Researchers have introduced ATOM-Bench, a new real-world benchmark designed to evaluate the atomic skills and compositional generalization capabilities of robotic manipulation policies. The benchmark includes 30 atomic tasks and 24 held-out compositional tasks, utilizing 3,000 human demonstrations for fine-tuning and evaluation. Initial tests on five representative policies revealed that while current models can grasp basic instruction-grounding, they struggle with fine-grained motor skills and reliably composing learned skills for novel tasks. AI
IMPACT This benchmark aims to improve the real-world generalization of robotic manipulation policies, addressing a key challenge in AI for robotics.