Event

EAST: Digital Twin Fault Injection Testing for Embodied AI Safety

1 reports · 1 sources

Yuhao Tian and other authors propose EAST (Embodied AI Safety Testing), a digital twin framework that translates literature-grounded safety evidence into fault injection tests covering perception, instructions, environment, actions, and timing, and is equipped with an expert-reviewed safety fault taxonomy and eight executable operators. Its process has three steps: generating test specifications, verifying during simulated execution whether faults actually occur, and independently evaluating safety challenges, violations, task outcomes, and behavioral characteristics using simulator state evidence. The authors compare it with alternative testing strategies under a matched budget on six RoboTwin manipulation tasks and two policy architectures, ACT and RDT. According to the authors' report, across five rounds of comparison EAST covered all 74 task-policy-operator combinations, of which 63 obtained valid evidence, and the safety challenge exposure rate was close to that of the Random-Legal and Rule-Based methods.