scaling-rl

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.