A new robot architecture called CoreSense checks whether a machine's memories of past failures are still trustworthy before letting it act.
Researchers built a system that gates a robot's recall of prior failures through checks on scope, timing, contradiction, and evidence quality before it decides to proceed, ask for another look, hold off, or flag a human. Tested on two benchmarks, CableTrace-120 and BotFails-200, the gate cut unsafe proceed rates from 20% and 40% down to 0%. A separate raw-video-only policy also hit 0% unsafe proceeds, but it overblocked every single normal episode, essentially refusing to act at all. Other tests on real robot telemetry and a cloud deployment produced mixed results, including one negative finding on transferring grip-loss detection.
The paper is upfront that this is an auditability pattern, not proof a robot can safely fix its own mistakes, and not a certified safety system. That distinction matters because zero unsafe proceeds sounds like a win until you notice the all-blocking baseline hits the same number by refusing to do anything useful at all.
Call it a lie detector for robot memory: good for catching bad recall, not a substitute for deciding whether the robot should be trusted with the task in the first place.