safe-policy-learning

The paper introduces Safe Contrastive Reinforcement Learning (Safe-CRL) to address overestimation bias in goal-conditioned policy learning under failure termination, using one-bit failure signals to enhance safety and performance in robot navigation and locomotion tasks.

This paper introduces Opal (Opportunity-aware Policy Authorization for Laboratories), a framework that certifies whether adaptive experimentation should be enabled by precommitting to non-trivial adaptation, controlled target risk, and positive executed value after cost. It establishes an impossibility boundary and demonstrates the method on a Cell Painting dataset, achieving risk control and positive value.