Researchers introduced GUARD, a method for steering generative robot policies during execution without changing the pretrained policy. It combines task rewards with anomaly scores—signals that an action falls outside patterns supported by training data—and uses a learned Q-function to guide the policy’s action-generation process.
In four simulated dexterous-manipulation tasks, the authors report average success rates of 96% under normal conditions and 80% under perturbed dynamics, beating the strongest baseline by 21 and 20 percentage points. They also report success-rate gains of 40 points for in-hand screwdriver rotation and 25 points for reorientation in direct sim-to-real tests.
Why it matters
For robotics teams adapting generative policies to changing conditions, GUARD offers a way to use training-data anomalies as a warning signal during action selection; the paper reports promising results, but does not establish how well the method works beyond its tested tasks.
