A new research paper introduced on Hugging Face presents ACG-Bench, a benchmark for evaluating dual-arm vision-language-action compositional generalization, and proposes AE-VLA—a policy combining skill-specific LoRA adapters with arm-wise attention on a π_0.5 backbone.
The approach achieves 21.53% simulation success on unseen object-task compositions, compared to just 2.94% for a single π_0.5 policy and 5.53% for two independently controlled π_0.5 policies. That’s a 7x improvement over the best prior method.
On physical Stretch RE1-2 robots, AE-VLA hits 39% mean success across five unseen conditions, up from 10% for the strongest baseline. The research addresses a fundamental challenge in robotics: teaching robots to generalize to novel combinations of skills and objects they haven’t encountered during training.
The key innovation is the combination of skill-specific LoRA adapters—which allow the model to learn specialized behaviors for different manipulation tasks—and arm-wise attention mechanisms that enable coordinated bimanual control. The π_0.5 foundation model provides robust visual reasoning, while the modular adapters enable rapid adaptation to new tasks.
Compositional generalization remains one of the hardest problems in embodied AI. Standard policies fail when encountering new object-task pairs because they learn rigid associations rather than composable primitives. AE-VLA’s modular architecture lets the system recombine learned skills in novel ways.
The benchmark and model are open-sourced on GitHub, providing the community with a standardized way to measure progress on bimanual manipulation generalization—a capability critical for practical robot deployment in warehouses, manufacturing, and domestic settings.