Data and benchmarks
A Fair Test of Dual-Arm Skill Recomposition
Use unseen combinations, explicit coordination constraints, controlled policy comparisons, and separate simulation and hardware reporting to assess compositional behavior.
A benchmark can test recomposition by giving both arms familiar atomic skills in combinations the policy has not seen, then requiring the overall goal, physical milestones, and specified order or timing constraints to be satisfied. The test asks whether the policy can coordinate those familiar skills in a new arrangement.
Its authors classify 6 conditions as in-domain and 17 as unseen compositions involving reordering, synchronization, both together, or composition across tasks. Every method receives the same per-arm atomic prompts, making the new coordination requirement the central variable under examination.
Success is not defined solely by reaching the final task goal. A policy must also complete physical milestones and obey the specified order or timing constraints. These criteria distinguish a valid coordinated execution from one that reaches an outcome without following the tested composition.
The authors’ architectural study examines arm-token grouping, skill-specific LoRA adapters called SkillLoRA, and arm-wise attention, then combines selected choices in a system named AE-VLA.
Simulation and hardware results should be reported separately because they come from different settings. The authors report 21.53% generalization success for AE-VLA in simulation, compared with 2.94% for Single π₀.₅, 3.06% for MA-VLA, and 5.53% for two independently controlled π₀.₅ policies. On physical SO101 robots, they report 39.00% mean success across five unseen conditions, versus 10.00% for the strongest baseline. These figures are author-reported results from the primary source, not independent confirmation, but they illustrate how held-out compositions, constraint-aware scoring, controlled comparisons, and separate physical evaluation can make recombination claims testable.