VA-Bench Finds a Gap Between Spatial Understanding and Robot Execution
The VA-Bench paper evaluates 12 multimodal models in closed-loop robot tasks spanning observation, reasoning, action, and revision. The source reports that leading models achieved near-perfect target localization and action-semantic scores, but complete-task success remained around half, with online correction, fine-grained control, active perception, and bimanual coordination as major weaknesses.
The benchmark highlights that reliable embodied spatial intelligence requires more than recognizing targets or understanding task instructions.
机器之心 2026-10-05 11:43 辽宁
VA-Bench 以“观察—推理—行动—修正”的闭环测试 12 个多模态模型：目标识别与操作语义接近满分，但完整任务成功率仅约一半，主动观察、精细控制和在线修正仍是主要瓶颈。
大模型已经越来越会「看懂」空间：它能判断物体在机械臂哪一侧，也能识别该操作哪个目标、理解任务怎么做。但如果真的把机械臂交给它，问题会变得更具体：该移动多少毫米？当前视角够不够？抓空以后，又该怎么调整？