Estimate, Don't Imitate: Reusing Differentiable State-Based Policies for Visuomotor Control
A visual state estimator is trained using both direct state supervision and an action-consistency loss backpropagated through the frozen, differentiable expert, which consistently outperforms direct pixel-to-action imitation from the same expert demonstration corpus.
Denis Shcherba, Adrian Abel, Eckart Cobo-Briesewitz et al.
· 0 citations