Step 3: Distill to touch.
A tactile state estimator learns to map real tactile signals to the privileged sensor teacher's latent. At deploy time the cameras are unnecessary: the policy runs on tactile + proprioception alone. This results in robust real-world deployable policies