Abstract
Industrial pick-and-place lines demand near-zero failure tolerance, yet many safety layers still wait for a fault to manifest before they trigger an indiscriminate Emergency Stop. We ask whether latent prediction error can instead serve as a graded recovery signal. Our closed loop combines a frozen DINOv2 encoder, an RSSM whose KL-divergence prediction error gives us an online risk reading, an ANFIS risk mapper, and a controller that applies bounded trajectory corrections scaled by that risk coefficient. In Webots with a UR5e arm and condition-based TRANSIT dynamics, we evaluate four recoverable motion disturbances (N = 40 episodes per mode per scenario) against a binary fail-safe halt (OFF), and we treat two semantic or actuation faults as outside joint-space recovery. RSR gains concentrate on the harder disturbances, Jerky (+25.0 pp) and Speed (+42.5 pp, with non-overlapping Wilson 95% CIs), while smoothing gains concentrate on Vibration, Jerky, and Pause, where mean joint deviation drops by 15.4–44.8% (Pause peaks at −44.8%); Speed's smoothing stays flat under the same joint-velocity cap. OFF/ON is a deployed-policy comparison. CPU-only latency is P95 = 34.24 ms, which sits 65.8% below the 100 ms design budget.