GUARD steers robot hand policies away from risky actions without changing the policy

Researchers led by Aryaman Gupta have presented GUARD, a method that steers generative robot policies away from failures by combining sparse task rewards with anomaly scores learned from the policy's training data, in a paper posted on arXiv.

A learned Q-function turns both signals into gradients that guide the denoising process toward actions that raise the task return while keeping trajectories in well-covered regions, without changing the pretrained policy.

On four Adroit dexterous manipulation tasks, GUARD averaged 96% success under normal dynamics and 80% under perturbed dynamics, 21 and 20 points above the best baseline. In real-world tests it raised success rates by 40 and 25 points on in-hand screwdriver rotation and reorientation.