Arm-wise Compositional Generalization in Dual-Arm Vision-Language-Action Models

Arm-wise Compositional Generalization in Dual-Arm Vision-Language-Action Models. By Zaibin Zhang and twelve others, from the Dalian University of Technology, the University of Oxford, and three other institutions. Published on the fifth of October. No code repository is linked to it.

Generalization in multi-arm collaboration can be studied as composition: taking atomic skills that are already familiar and combining them across arms in new ways. Existing evaluations, though, give limited insight into which training and architectural choices support that ability, and into how the answer shifts with the kind of coordination a task requires. A policy might handle each arm's skill perfectly and still fail when those skills have to happen in a new order, or at the same moment. Without a shared test that separates those cases, comparisons between methods say little about why one generalizes and another does not.

ACG-Bench, for Arm-wise Compositional Generalization, is a common testbed for studying how dual-arm policies recompose skills. Its unseen compositions cover four kinds of novelty: reordering, synchronization, the two combined, and composition across tasks. Success is strict. A run counts only if it achieves the task goal, satisfies physical milestones, and respects the specified order or timing constraints. Reaching the right end state by the wrong route does not count, which keeps the measure about composition rather than luck.

The benchmark holds twenty-three task-condition pairs across eight task families, made up of six in-domain conditions and seventeen unseen compositions. Every method receives the same atomic prompt for each arm. The study uses pi zero point five as a common vision-language-action backbone, and compares representative strategies, some based on data augmentation and some architectural, with shared source data and one evaluation protocol. Holding the backbone, the data and the prompts fixed is what lets differences in results be traced back to the strategy itself.

The architectural study looks at three choices: grouping tokens by arm, skill-specific LoRA adapters called SkillLoRA, and arm-wise attention, or AWA. What it highlights is that skill-conditioned parameters and attention structure complement each other.

Combining these choices yields AE-VLA. In simulation it achieves twenty-one point five three percent generalization success. For comparison, a single pi zero point five policy reaches two point nine four percent, MA-VLA reaches three point zero six percent, and two independently controlled pi zero point five policies reach five point five three percent.

On physical SO101 robots, AE-VLA reaches thirty-nine percent mean success across five unseen conditions, against ten percent for the strongest baseline. The ordering seen in simulation holds on hardware, with the gap between them still wide.

Even the best configuration succeeds on roughly one unseen composition in five in simulation, which shows how much headroom the benchmark leaves. The authors present the results as empirical guidance for designing dual-arm policies that generalize beyond fixed training routines, and the guidance is specific: parameters conditioned on skills and attention organised by arm do more together than either does alone.