Action Self-Distillation: One VLM Trains Itself Into a Generalist Navigation Policy
Loading video
Loading videoAction Self-Distillation splits one VLM into a privileged self-teacher that sees demonstrations and language labels and an unprivileged student that sees only images. From about 100 hours of robot driving data, the student learns language goal-conditioned reasoning and navigation with no human-annotated labels.
Category: humanoid
Author: @BowenFengbf
Date: 2026-10-22T00:00:00
Duration: 59.966s
Reference: https://arxiv.org/abs/2610.17593
