When to Act, Ask, or Learn: Uncertainty-Aware Policy Steering

Researchers from UC Berkeley have developed a novel framework called Uncertainty-Aware Policy Steering (UPS) that addresses a critical challenge in robotics: how autonomous systems should handle uncertainty during real-world deployment. The system enables robots to intelligently choose between three strategies—executing an action, asking for task clarification via natural language, or requesting human intervention—based on calibrated confidence levels. This approach moves beyond existing policy steering methods that often assume perfect verifier calibration, instead acknowledging that Vision-Language Models (VLMs) like GPT-4V can be overconfident in their judgments, leading to degraded performance when task specifications are ambiguous or when low-level policies are incapable.

The technical innovation lies in using conformal prediction to statistically calibrate the composition of VLMs and pre-trained base policies (like diffusion policies), providing formal assurances about when the system should act versus ask for help. After collecting interventions during deployment, UPS employs residual learning to improve the base policy's capabilities, creating a continuous learning loop with minimal human feedback. Experiments in simulation and on physical hardware show UPS can effectively disentangle different uncertainty types and reduces expensive human interventions by approximately 40% compared to uncalibrated baselines, marking significant progress toward more autonomous, adaptable robotic systems that know their limits.