SO-101 Bench: Accessible Spatial Manipulation

Isaac Lab digital-twin benchmark for language-conditioned spatial tasks on the low-cost SO-101 arm (138 objects, 916 episodes); sim ranks policies exactly as the real arm. Best VLAs ~61%, an LLM-agent harness 83.2%.基于 Isaac Lab 数字孪生的低成本 SO-101 机械臂语言条件空间操作基准（138 个物体、916 个评测回合）；仿真中的策略排名与真机完全一致。最好的 VLA 约 61%，由 LLM 写代码调用抓取策略的方案达 83.2%。

@misc{hickok2026so,
  title = {{SO-101 Bench: Accessible Spatial Manipulation}},
  author = {Truman Hickok},
  year = {2026},
  eprint = {2610.36420},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2610.36420},
}