Latent Policy Steering

Researchers introduced Latent Policy Steering (LPS), a new offline reinforcement learning method that directly integrates Q-gradients from the original action space into latent-space optimization, ach…