Steers frozen robot foundation models including VLAs, diffusion policies, and world models using human corrections via action inversion in latent space without fine-tuning the base policy. Outperforms supervised fine-tuning and latent RL with only 5–20 human interventions.