← Back to Robotics
cs.RO

A cheap, copyable test bench for robot learning models

Alex S. Huang, Jiahui Zhang, Shiqing Tang, Yu Xiang

May 20, 2026

Evaluating robot learning models in the real world has been blocked by expensive, hard-to-replicate benchmarks. VLA-REPLICA solves this with a low-cost setup built from standard components that any lab can assemble, complete with a diverse task suite and adaptation data. Tests on imitation learning and state-of-the-art VLA models show consistent results across independently built copies, proving the benchmark actually works as intended.
Published as VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models arXiv:2605.20774
Read the original paper →