做机器人 VLA 部署的团队注意了——不同架构的失败模式完全不同，用错监控等于白费功夫。建议直接看方向反转率这个通用指标，并试试 SafeContract 工具包。
研究发现视觉-语言-动作（VLA）模型在电机指令层面存在根本性、可预测的失败差异。通过对 VQ-BeT、Diffusion Policy 和 ACT 三种架构在 PushT 和 ALOHA 14-DOF 双臂操作任务上进行 450 次评估，发现方向反转率是通用失败预测指标（AUROC 最高 0.93），而急动度监控仅对离散令牌架构有效，速度监控在连续架构中几乎无效（AUROC 仅 0.41-0.52）。研究强调架构匹配的监控选择至关重要，并开源了 SafeContract 工具包。
How VLAs Fail Differently: Black-Box Action Monitoring Reveals Architecture-Specific Failure Signatures
We discover that VLA architectures fail in fundamentally different, predictable ways at the motor-command level. Running VQ-BeT, Diffusion Policy, and ACT on identical evaluation protocols (n=450 episodes across PushT and ALOHA 14-DOF bimanual manipulation), we find: (1) direction reversal rate is a universal failure predictor across all three architectures (AUROC=0.93, 0.79, 0.91; p<0.001); (2) jerk monitoring is predictive only for discrete-token architectures, following a discrete-to-continuous gradient (0.88, 0.69, 0.41); (3) velocity violations alone are non-predictive everywhere (AUROC 0.41-0.69), yet velocity checking is the most common safety mechanism in VLA deployment code; and (4) for continuous-family VLAs, velocity monitoring provides effectively zero predictive signal (AUROC=0.52 on ACT, 0.41 on Diffusion), proving that architecture-matched monitor selection is essential. These results quantify a monitoring consequence of the well-known discrete/continuous VLA distinction: the two families produce qualitatively different failure signatures that require different monitors. No single monitor works universally; architecture-matched selection is required. This finding was enabled by SafeContract, a training-free, black-box action monitoring toolkit with conformal calibration. Code: https://github.com/krishnam94/vla-edge